Speech-to-text evidence benchmark
Cross-benchmark speech recognition, release cohorts, local runtimes and documented API rates.
172 model configurations · 975 source rows · 11 accuracy protocols · 464 throughput measurements · checked 2026-10-19
Rank uses distinct English short-form, CORAAL and five-language FLEURS families. WER columns are lower-is-better; the percentile index is higher-is-better. Two mature families permit a provisional index.
35 models · 3 mature core families · default coverage: 2+ primary families
| Model / configuration | Rank index | English · 7 sets ↓ | CORAAL / long-form ↓ | FLEURS · 5 languages ↓ | Download / params | API / audio hour |
|---|---|---|---|---|---|---|
| universal-3-pro | 85.2 universal-3-pro · provisional rank Mean percentile across 2 mature core families: FLEURS · 5 languages, CORAAL / long-form. | — | 12.83† | 2.44† | — | — |
| scribe_v2 | 67.6 scribe_v2 · provisional rank Mean percentile across 3 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages, CORAAL / long-form.
| 4.06† | 10.67† | 2.12† | — | — |
| omniASR-LLM-7B-v2 | 64.6 omniASR-LLM-7B-v2 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 6.62† | — | 3.66† | — | Self-hosted |
| Voxtral-Small-24B-2507 | 64.3 Voxtral-Small-24B-2507 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 5.21† | — | 3.01† | 24,261.8M params | Self-hosted |
| distil-large-v3.5 | 63.3 distil-large-v3.5 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, CORAAL / long-form.
| 5.66† | 17.61† | — | 756.4M params | Self-hosted |
| canary-1b-v2 | 62.9 canary-1b-v2 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 5.81† | — | 3.38† | 979M params | Self-hosted |
| enhanced | 62.3 enhanced · provisional rank Mean percentile across 2 mature core families: FLEURS · 5 languages, CORAAL / long-form. | — | 14.29† | 3.58† | — | — |
| gpt-transcribe | 62.1 gpt-transcribe · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 4.70† | — | 2.41† | — | — |
| whisper-large-v3 | 61.3 whisper-large-v3 · provisional rank Mean percentile across 3 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages, CORAAL / long-form.
| 6.04† | 18.89† | 3.22† | 1,543.5M params | Self-hosted |
| Voxtral-Mini-3B-2507 | 60.9 Voxtral-Mini-3B-2507 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 5.63† | — | 3.35† | 4,676.3M params | Self-hosted |
| gemma-4-E4B-it | 60.2 gemma-4-E4B-it · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 9.98† | — | 4.94† | 7,996.2M params | Self-hosted |
| resonant-1-flash | 59.8 resonant-1-flash · provisional rank Mean percentile across 3 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages, CORAAL / long-form.
| 4.16† | 12.74† | 2.69† | — | — |
| Whisper large-v3-turbo | 59.0 Whisper large-v3-turbo · provisional rank Mean percentile across 3 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages, CORAAL / long-form.
| 6.58† | 17.87† | 3.67† | 1,618 MB | Self-hosted |
| omniASR-CTC-7B-v2 | 58.0 omniASR-CTC-7B-v2 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 9.35† | — | 5.04† | — | Self-hosted |
| universal-3-5-pro | 56.8 universal-3-5-pro · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 4.46† | — | 2.17† | — | $0.21† |
| resonant-1 | 56.6 resonant-1 · provisional rank Mean percentile across 3 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages, CORAAL / long-form.
| 4.15† | 12.87† | 2.69† | — | — |
| azure-speech-06-2026 | 56.1 azure-speech-06-2026 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 4.24† | — | 1.99† | — | — |
| Phi-4-multimodal-instruct | 55.6 Phi-4-multimodal-instruct · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 5.10† | — | 3.32† | 5,574.5M params | Self-hosted |
| cohere-transcribe-03-2026 | 55.6 cohere-transcribe-03-2026 · provisional rank Mean percentile across 3 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages, CORAAL / long-form.
| 4.85† | 15.34† | 3.50† | 2,065.8M params | Self-hosted |
| parakeet-ctc-1.1b | 54.7 parakeet-ctc-1.1b · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, CORAAL / long-form.
| 6.02† | 23.91† | — | 1,062.6M params | Self-hosted |
| symphony | 53.6 symphony · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 4.47† | — | 2.63† | — | — |
| Voxtral Mini 4B Realtime 2602 | 53.2 Voxtral Mini 4B Realtime 2602 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 6.12† | — | 4.66† | — | Self-hosted |
| azure-speech-07-2026 | 52.5 azure-speech-07-2026 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 3.93† | — | 1.80† | — | — |
| Parakeet TDT 0.6B v3 | 50.5 Parakeet TDT 0.6B v3 · provisional rank Mean percentile across 3 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages, CORAAL / long-form.
| 4.94† | 15.57† | 3.79† | 2,508 MB | Self-hosted |
| thewhisper-large-v3-turbo | 50.3 thewhisper-large-v3-turbo · provisional rank Mean percentile across 3 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages, CORAAL / long-form.
| 4.66† | 17.95† | 2.95† | 808.9M params | Self-hosted |
| Nemotron 3.5 ASR Streaming 0.6B | 49.6 Nemotron 3.5 ASR Streaming 0.6B · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 7.96† | — | 6.63† | 2,368 MB | Self-hosted |
| parakeet-tdt-0.6b-v2 | 48.8 parakeet-tdt-0.6b-v2 · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, CORAAL / long-form.
| 4.82† | 17.64† | — | — | Self-hosted |
| muse-voice-transcribe | 43.6 muse-voice-transcribe · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 4.93† | — | 3.73† | — | — |
| Qwen3-ASR-1.7B-hf | 40.2 Qwen3-ASR-1.7B-hf · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 4.43† | — | 3.31† | — | Self-hosted |
| VibeVoice-ASR-HF | 38.2 VibeVoice-ASR-HF · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 5.69† | — | 6.78† | 8,330.3M params | Self-hosted |
| Qwen3-ASR-0.6B-hf | 35.6 Qwen3-ASR-0.6B-hf · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 5.16† | — | 5.51† | — | Self-hosted |
| canary-qwen-2.5b | 33.3 canary-qwen-2.5b · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, CORAAL / long-form.
| 4.55† | 19.05† | — | 2,559.4M params | Self-hosted |
| MOSS-Transcribe-Diarize | 26.8 MOSS-Transcribe-Diarize · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 4.75† | — | 5.51† | 908.5M params | Self-hosted |
| multilingual | 20.6 multilingual · provisional rank Mean percentile across 2 mature core families: Open ASR · October 17, 2026, FLEURS · 5 languages.
| 3.92† | — | 3.78† | — | — |
| parakeet-tdt-0.6b-v3 (fast-gpu-asr) | Unranked | 4.82† | — | 3.74† | — | Self-hosted |
◊ provisional index: hover or click for reasons and sources · † qualified evidence / current service rate · — no matched observation. Small models and New research expose limited coverage. Quantized GGUF configurations keep separate identities. MB is download size; M params is parameter count; neither is runtime RAM. Unknown quantities are excluded by size filters.
Local throughput & load time
Audio seconds per second of wall time (× realtime), higher is faster. Compare the same workload, hardware, backend and batch size. These are tiny runtime checks, not accuracy families or streaming latency measurements. Model loading is excluded from throughput and shown separately when reported.
3 measurements on the selected workload
| Model / configuration | × realtime | Hardware / runtime | Batch | Workload / sample | Model load | Source / run date |
|---|---|---|---|---|---|---|
| Phonon-2 | 174.00 | Apple M5 MacBook Air MLX, Apple GPU | 1 | Fermion 20 dictations | — | Source ↗ Run date unspecified |
| Parakeet TDT 0.6B v3 | 104.90 | Apple M5 MacBook Air FluidAudio, Core ML | 1 | Fermion 20 dictations | — | Source ↗ Run date unspecified |
| Parakeet Redux | 27.80 | Apple M5 MacBook Air FluidAudio, Core ML | 1 | Fermion 20 dictations | — | Source ↗ Run date unspecified |
End-of-utterance response latency
Moonshine paper Table 2: elapsed time between the end of speech and the returned transcript on an Apple MacBook M3. The source does not identify a p50/p95 statistic or a pinned utterance set. This is separate from time to first token, model loading and throughput. All rows were measured by Moonshine; comparisons to other developers’ models are competitor evidence.
| Model | Response ms ↓ | Hardware / statistic | Operator / source |
|---|---|---|---|
| moonshine-tiny | 27 | Apple MacBook M3 Statistic unspecified | Moonshine AI · author_reported · Paper Table 2 ↗ |
| moonshine-base | 44 | Apple MacBook M3 Statistic unspecified | Moonshine AI · author_reported · Paper Table 2 ↗ |
| Moonshine v2 Tiny | 50 | Apple MacBook M3 Statistic unspecified | Moonshine AI · author_reported · Paper Table 2 ↗ |
| Moonshine v2 Small | 148 | Apple MacBook M3 Statistic unspecified | Moonshine AI · author_reported · Paper Table 2 ↗ |
| Moonshine v2 Medium | 258 | Apple MacBook M3 Statistic unspecified | Moonshine AI · author_reported · Paper Table 2 ↗ |
| whisper-tiny | 289 | Apple MacBook M3 Statistic unspecified | Moonshine AI · vendor_comparison · Paper Table 2 ↗ |
| whisper-base | 553 | Apple MacBook M3 Statistic unspecified | Moonshine AI · vendor_comparison · Paper Table 2 ↗ |
| whisper-small | 1940 | Apple MacBook M3 Statistic unspecified | Moonshine AI · vendor_comparison · Paper Table 2 ↗ |
| whisper-large-v3 | 11286 | Apple MacBook M3 Statistic unspecified | Moonshine AI · vendor_comparison · Paper Table 2 ↗ |
Published audio tariffs
Each rate retains provider, currency, tier, mode and date. Minute tariffs are normalized by multiplying by 60. EUR stays EUR. Current service rates do not guarantee the exact model endpoint used in a benchmark. Optional features, volume tiers and promotions affect the bill; see the source configuration.
| Model / service | Per audio hour | Published rate | Provider / mode / tier | Scope & source |
|---|---|---|---|---|
| Deepgram Nova-3 Monolingual | USD 0.258 | 0.0043 / min | Deepgram · pre-recorded Pay As You Go · 2026-10-19 | Original per-minute rate × 60. Add-ons, Growth discounts and taxes excluded; no accuracy inherited from another configuration. Rate card ↗ |
| Deepgram Nova-3 Monolingual | USD 0.288 | 0.0048 / min | Deepgram · streaming-current Pay As You Go · 2026-10-19 | Original per-minute rate × 60. Current promotional streaming rate, not the regular rate. Add-ons, Growth discounts and taxes excluded; no accuracy inherited from another configuration. Rate card ↗ |
| Deepgram Nova-3 Multilingual | USD 0.312 | 0.0052 / min | Deepgram · pre-recorded Pay As You Go · 2026-10-19 | Original per-minute rate × 60. Add-ons, Growth discounts and taxes excluded; no accuracy inherited from another configuration. Rate card ↗ |
| Deepgram Nova-3 Multilingual | USD 0.348 | 0.0058 / min | Deepgram · streaming-current Pay As You Go · 2026-10-19 | Original per-minute rate × 60. Current promotional streaming rate, not the regular rate. Add-ons, Growth discounts and taxes excluded; no accuracy inherited from another configuration. Rate card ↗ |
| universal-3-5-pro | USD 0.21 | 0.21 / hour | AssemblyAI · pre-recorded Pay As You Go · 2026-10-19 | Current service generation/tariff. Optional add-ons and realtime API products excluded; historic benchmark endpoint is not guaranteed. Rate card ↗ |
| AssemblyAI Universal-2 | USD 0.15 | 0.15 / hour | AssemblyAI · pre-recorded Pay As You Go · 2026-10-19 | Current service generation/tariff. Optional add-ons and realtime API products excluded; historic benchmark endpoint is not guaranteed. Rate card ↗ |
| solaria-3 | EUR 0.16 | 0.16 / hour | Gladia · async Self-serve; 0–100 hours/month · 2026-10-19 | EUR retained as published. No currency conversion. Optional add-ons and volume discounts excluded; real-time rate describes the current service rather than a pinned benchmark model revision. Rate card ↗ |
| solaria-3 | EUR 0.68 | 0.68 / hour | Gladia · real-time Self-serve; 0–100 hours/month · 2026-10-19 | EUR retained as published. No currency conversion. Optional add-ons and volume discounts excluded; real-time rate describes the current service rather than a pinned benchmark model revision. Rate card ↗ |
| Voxtral Mini Transcribe 2602 API | USD 0.18 | 0.003 / min | Mistral AI · transcription Standard API · 2026-10-19 | Published USD 0.003/minute × 60. No open-weight Realtime score is assigned to this hosted transcription generation. Rate card ↗ |
Audio and text token tariffs
USD per million tokens. Input audio, input text and output text are different meters. Soniox’s published hourly equivalents depend on a token workload; they are not inserted as flat hourly tariffs. These current service prices do not assign an async benchmark score to a streaming endpoint.
| Service | Input audio / 1M | Input text / 1M | Output text / 1M | Mode / date / source |
|---|---|---|---|---|
| Soniox current service | $1.5 | $3.5 | $3.5 | async · 2026-10-19 · Token rate card ↗ |
| Soniox current service | $2 | $4 | $4 | streaming · 2026-10-19 · Token rate card ↗ |
Self-hosted means deployment-specific compute cost, not zero. No RAM, energy, hosting cost or audio-hour price is inferred from a model download, parameter count, throughput or token rate.
Methodology, provenance & sources
This is an index of published speech recognition evidence. We have not executed the transcription models. WER is 100 × (substitutions + deletions + insertions) / reference words, so it can exceed 100%. Punctuation, diarization, hallucinations and latency need their own measurements.
The core matrix selects three distinct corpus families: the fixed seven-set English short-form panel, CORAAL long-form interviews, and a fixed French/German/Spanish/Italian/Portuguese FLEURS panel. All named components are required for a panel mean. Language CSVs retain separate pinned revisions; no missing language or dataset is filled from another variant. CORAAL alone supplies the long-form core value so Earnings22 is not counted twice. MCV, MLS, extra languages, noise/Monsoon columns and author/runtime variants remain secondary. Hindi Monsoon uses lattice-reference OI-WER with the voi-oiwer scorer; it is not inserted as standard WER or a private Hindi average.
Index is the equal-weight mean of within-family percentile ranks, with midranks for ties. A mature family requires six scored configurations from three organizations. Only one runtime variant for the same weights identity enters a reference cohort, and one selected variant per family contributes. Two mature families permit a provisional index; three remove only the limited-coverage qualification. A lone family is unranked. The default matrix requires two primary families for display; Small models, New research and Secondary expose narrower evidence.
Unknown execution operators, author/competitor/community/imported evidence, conflicting observations, and unknown/unreleased availability qualify an index. Cohort qualifications propagate to every dependent percentile before search or display filters. ◊ exposes actual reasons and every contributing component source on hover or keyboard-accessible click. Publication by a maintainer and source verification do not establish independent reproduction.
The release slider uses verified availability dates, not paper publication, result dates or Hub repository creation. Unknown releases appear only in the current view. Quantized runtime conversions have separate identities and do not inherit upstream release dates, scores, size or memory use. Catalog confidence intervals and sample counts remain in the downloadable data. Legacy-published catalog cells can lack a run date or engine SHA.
Source preference within a matched protocol is maintainer publication, independent third party, community, model author, then vendor/secondary evidence. Alternative observations are preserved; conflicts are not averaged. Assumed, incompatible and private/vendor-internal values cannot supply core coverage, maturity or index. FFASR’s private corpus is not assigned a core vote. The original Phonon-2 164 MB package and its approximately 740 MB GGUF Q8_0 artifact are different configurations.
Further gaps include matched diarization and silence/hallucination tests, broader independently verified release dates, and comparable streaming p50/p95 latency and peak RAM. The latency paper and tiny runtime speed clips do not supply these missing measurements.
Download the source data JSON · 198 documented sources. Prices retain their own as-of dates in historical model views.
Open ASR English short-form results, October 17, 2026
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Maintainer publication. The CSV does not attribute execution for each row; publication is not proof of independent reproduction. Per-dataset WER columns are imported verbatim. Its avg and RTFx aggregate columns are deliberately not imported; the dashboard default dataset mix has changed.
Open ASR dataset and metric documentation
Publication: benchmark_maintainer · execution: not_applicable · checked 2026-10-19. Documents cleaned reference variants, session-level Earnings22 scoring and the distinction between URGENT24 clean and noisy versions. Actual dataset revisions and individual run operators are not pinned by the CSV.
Open ASR Phonon-2 evaluation configuration
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Configuration names librispeech test.clean/test.other; ami_cleaned, gigaspeech_cleaned, voxpopuli_cleaned_aa, spgispeech and URGENT test splits; Earnings22 uses ArtificialAnalysis/Earnings22-Cleaned-AA-chunked. Documents H200, batch 64, bfloat16 and full sets. This is documentation, not proof that a particular CSV row used that exact script revision.
Open ASR Phonon-2 reference evaluation engine
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Expands Phonon weights to a dense Transformers graph. Its CUDA throughput is not the packed Apple runtime throughput. Do not equate leaderboard H200 batch speed with Mac single-stream measurements.
Fermion Research: Introducing Phonon-2, Tables 1–3
Publication: model_author · execution: mixed · checked 2026-10-19. Unmarked Phonon-2 and Phonon-1 accuracy rows are author runs; Redux is a competitor run by Fermion. Daggered accuracy rows are copied from an unspecified leaderboard revision, not new Fermion runs. Short dataset labels do not pin cleaned-reference versions or dataset commits. Keep this table separate from the current maintainer protocol. Table 3 is a 20-dictation throughput workload, not a standardized accuracy family.
Fermion Research model release changelog
Publication: model_author · execution: not_applicable · checked 2026-10-19. Explicit weights release entries: Phonon-2 on September 29, 2026; Phonon-1 on August 28, 2026. Core ML is a separate October 4 package release.
FermionResearch/Phonon-2 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Low-bit Parakeet TDT 0.6B v3 derivative; English. Download size is not runtime memory.
FermionResearch/Phonon-1 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Low-bit Qwen3-ASR derivative; English. The Phonon-2 launch comparison is a different protocol from the eight-set Phonon-1 model card.
moondream/parakeet-redux model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Low-bit Parakeet derivative. Only the Fermion launch comparison is imported; exact launch checkpoint revision was not reported.
nvidia/canary-180m-flash model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Launch-table accuracy is a leaderboard row quoted by Fermion; no matching current maintainer row is imported.
abr-ai/niagara-9m-batch.en model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact public non-causal batch.en model; the commercial streaming model is a different deployment.
abr-ai/niagara-19m-batch.en model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact public non-causal batch.en model; the commercial streaming model is a different deployment.
abr-ai/niagara-38m-batch.en model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact public non-causal batch.en model; the commercial streaming model is a different deployment.
abr-ai/niagara-84m-batch.en model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact public non-causal batch.en model; the commercial streaming model is a different deployment.
nvidia/parakeet-tdt-0.6b-v3 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Standard leaderboard configuration; fast-gpu-asr reruns are separate configurations and are not substituted.
openai/whisper-large-v3-turbo model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Open-weight large-v3-turbo only; no hosted Whisper or GPT transcription scores are inferred.
Qwen/Qwen3-ASR-0.6B-hf model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact Transformers-native -hf checkpoint from the leaderboard. Its original-model release date is not assigned to this variant.
Qwen/Qwen3-ASR-1.7B-hf model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact Transformers-native -hf checkpoint from the leaderboard. Its original-model release date is not assigned to this variant.
ibm-granite/granite-speech-5.0-470m-turboctc model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Apache-2.0 TurboCTC checkpoint; the -nc checkpoint is a distinct model and is not substituted.
mistralai/Voxtral-Mini-4B-Realtime-2602 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact 2602 open-weight generation. The launch article estimates its bytes from parameter count; that estimate is not stored as an observed download size.
nvidia/nemotron-3.5-asr-streaming-0.6b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact Nemotron 3.5 generation; offline short-form scoring does not establish streaming latency.
Edge0/ARK-ASR-0.6B model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: Edge0/ARK-ASR-0.6B. Model commit and evaluation date are unspecified unless the observation records them.
Edge0/ARK-ASR-0.6B Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
Edge0/ARK-ASR-3B model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: Edge0/ARK-ASR-3B. Model commit and evaluation date are unspecified unless the observation records them.
Edge0/ARK-ASR-3B Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
Edge0/Audio8-ASR-0.1B model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: Edge0/Audio8-ASR-0.1B. Model commit and evaluation date are unspecified unless the observation records them.
Edge0/Audio8-ASR-0.1B Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
bosonai/higgs-audio-v3-stt model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: bosonai/higgs-audio-v3-stt. Model commit and evaluation date are unspecified unless the observation records them.
bosonai/higgs-audio-v3-stt Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
CohereLabs/cohere-transcribe-03-2026 Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
distil-whisper/distil-large-v3.5 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: distil-whisper/distil-large-v3.5. Model commit and evaluation date are unspecified unless the observation records them.
distil-whisper/distil-large-v3.5 Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
espnet/owsm_ctc_v4_1B model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: espnet/owsm_ctc_v4_1B. Model commit and evaluation date are unspecified unless the observation records them.
FunAudioLLM/Fun-ASR-Nano-2512-hf model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: FunAudioLLM/Fun-ASR-Nano-2512-hf. Model commit and evaluation date are unspecified unless the observation records them.
FunAudioLLM/Fun-ASR-Nano-2512-hf Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
google/gemma-4-12B-it model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: google/gemma-4-12B-it. Model commit and evaluation date are unspecified unless the observation records them.
google/gemma-4-12B-it Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
google/gemma-4-E2B-it model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: google/gemma-4-E2B-it. Model commit and evaluation date are unspecified unless the observation records them.
google/gemma-4-E2B-it Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
google/gemma-4-E4B-it model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: google/gemma-4-E4B-it. Model commit and evaluation date are unspecified unless the observation records them.
google/gemma-4-E4B-it Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
HojoAI/Hojo-ASR-V1 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: HojoAI/Hojo-ASR-V1. Model commit and evaluation date are unspecified unless the observation records them.
ibm-granite/granite-4.0-1b-speech model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: ibm-granite/granite-4.0-1b-speech. Model commit and evaluation date are unspecified unless the observation records them.
ibm-granite/granite-4.0-1b-speech Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
ibm-granite/granite-speech-3.3-2b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: ibm-granite/granite-speech-3.3-2b. Model commit and evaluation date are unspecified unless the observation records them.
ibm-granite/granite-speech-3.3-2b Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
ibm-granite/granite-speech-3.3-8b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: ibm-granite/granite-speech-3.3-8b. Model commit and evaluation date are unspecified unless the observation records them.
ibm-granite/granite-speech-3.3-8b Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
ibm-granite/granite-speech-4.1-2b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: ibm-granite/granite-speech-4.1-2b. Model commit and evaluation date are unspecified unless the observation records them.
ibm-granite/granite-speech-4.1-2b Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
ibm-granite/granite-speech-4.1-2b-nar model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: ibm-granite/granite-speech-4.1-2b-nar. Model commit and evaluation date are unspecified unless the observation records them.
ibm-granite/granite-speech-4.1-2b-nar Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
ibm-granite/granite-speech-5.0-470m-turboctc-nc model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: ibm-granite/granite-speech-5.0-470m-turboctc-nc. Model commit and evaluation date are unspecified unless the observation records them.
ibm-granite/granite-speech-5.0-470m-turboctc-nc Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
microsoft/Phi-4-multimodal-instruct model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: microsoft/Phi-4-multimodal-instruct. Model commit and evaluation date are unspecified unless the observation records them.
microsoft/Phi-4-multimodal-instruct Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
microsoft/VibeVoice-ASR-HF model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: microsoft/VibeVoice-ASR-HF. Model commit and evaluation date are unspecified unless the observation records them.
microsoft/VibeVoice-ASR-HF Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
mistralai/Voxtral-Mini-3B-2507 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: mistralai/Voxtral-Mini-3B-2507. Model commit and evaluation date are unspecified unless the observation records them.
mistralai/Voxtral-Mini-3B-2507 Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
mistralai/Voxtral-Small-24B-2507 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: mistralai/Voxtral-Small-24B-2507. Model commit and evaluation date are unspecified unless the observation records them.
mistralai/Voxtral-Small-24B-2507 Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
nvidia/canary-1b-v2 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/canary-1b-v2. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/canary-1b-v2 Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
nvidia/canary-qwen-2.5b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/canary-qwen-2.5b. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/canary-qwen-2.5b Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
nvidia/nemotron-speech-streaming-en-0.6b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/nemotron-speech-streaming-en-0.6b. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/nemotron-speech-streaming-en-0.6b Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
nvidia/parakeet-ctc-1.1b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/parakeet-ctc-1.1b. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/parakeet-ctc-1.1b Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
nvidia/parakeet-tdt-0.6b-v2 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/parakeet-tdt-0.6b-v2. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-large-v3 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: openai/whisper-large-v3. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-large-v3 Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
OpenMOSS-Team/MOSS-Transcribe-Diarize model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: OpenMOSS-Team/MOSS-Transcribe-Diarize. Model commit and evaluation date are unspecified unless the observation records them.
OpenMOSS-Team/MOSS-Transcribe-Diarize Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
TheStageAI/thewhisper-large-v3-turbo model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: TheStageAI/thewhisper-large-v3-turbo. Model commit and evaluation date are unspecified unless the observation records them.
TheStageAI/thewhisper-large-v3-turbo Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
Open ASR long-form results, pinned October 17 snapshot
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. CORAAL mean is retained as published; it covers the entire CORAAL dataset per maintainer documentation. Earnings21/22 and TED-LIUM are detail columns; the overlapping full-table Average is not a separate core score. Negative RTFx sentinels are not imported.
nvidia/parakeet-ctc-0.6b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/parakeet-ctc-0.6b. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/parakeet-ctc-0.6b Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
nvidia/parakeet-rnnt-0.6b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/parakeet-rnnt-0.6b. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/parakeet-rnnt-0.6b Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
nvidia/parakeet-rnnt-1.1b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/parakeet-rnnt-1.1b. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/parakeet-rnnt-1.1b Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
nvidia/parakeet-tdt-1.1b model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/parakeet-tdt-1.1b. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/stt_en_conformer_ctc_large model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/stt_en_conformer_ctc_large. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/stt_en_conformer_ctc_small model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/stt_en_conformer_ctc_small. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/stt_en_conformer_transducer_large model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/stt_en_conformer_transducer_large. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/stt_en_fastconformer_ctc_large model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/stt_en_fastconformer_ctc_large. Model commit and evaluation date are unspecified unless the observation records them.
nvidia/stt_en_fastconformer_transducer_large model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/stt_en_fastconformer_transducer_large. Model commit and evaluation date are unspecified unless the observation records them.
Open ASR French results
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Exact language-specific WER columns, test split. Source labels do not establish model commits or individual execution operators. Hindi uses Voice Arena Monsoon, not FLEURS. Missing columns are never substituted.
Open ASR German results
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Exact language-specific WER columns, test split. Source labels do not establish model commits or individual execution operators. Hindi uses Voice Arena Monsoon, not FLEURS. Missing columns are never substituted.
Open ASR Spanish results
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Exact language-specific WER columns, test split. Source labels do not establish model commits or individual execution operators. Hindi uses Voice Arena Monsoon, not FLEURS. Missing columns are never substituted.
Open ASR Italian results
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Exact language-specific WER columns, test split. Source labels do not establish model commits or individual execution operators. Hindi uses Voice Arena Monsoon, not FLEURS. Missing columns are never substituted.
Open ASR Portuguese results
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Exact language-specific WER columns, test split. Source labels do not establish model commits or individual execution operators. Hindi uses Voice Arena Monsoon, not FLEURS. Missing columns are never substituted.
Open ASR Hindi results
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Public VoiceArena/Monsoon_hi_test, test split. Hindi uses lattice references and the voi-oiwer orthographic-variant scorer, not flattened-reference standard WER. The public Monsoon WER CSV column is preserved; no private Hindi average is inferred. Secondary evidence, outside core ranking.
Open ASR Dutch results
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Exact language-specific WER columns, test split. Source labels do not establish model commits or individual execution operators. Hindi uses Voice Arena Monsoon, not FLEURS. Missing columns are never substituted.
Open ASR Armenian results
Publication: benchmark_maintainer · execution: unspecified · checked 2026-10-19. Exact language-specific WER columns, test split. Source labels do not establish model commits or individual execution operators. Hindi uses Voice Arena Monsoon, not FLEURS. Missing columns are never substituted.
HojoAI/Hojo-ASR-Multi-V1 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: HojoAI/Hojo-ASR-Multi-V1. Model commit and evaluation date are unspecified unless the observation records them.
facebook/mms-1b-all model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: facebook/mms-1b-all. Model commit and evaluation date are unspecified unless the observation records them.
facebook/mms-1b-all Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
facebook/seamless-m4t-v2-large model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: facebook/seamless-m4t-v2-large. Model commit and evaluation date are unspecified unless the observation records them.
facebook/seamless-m4t-v2-large Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
nvidia/stt_hy_fastconformer_hybrid_large_pc model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: nvidia/stt_hy_fastconformer_hybrid_large_pc. Model commit and evaluation date are unspecified unless the observation records them.
bharatgenai/Shrutam-2 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: bharatgenai/Shrutam-2. Model commit and evaluation date are unspecified unless the observation records them.
bharatgenai/Shrutam-2 Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
Moonshine v2 paper, Tables 1–3
Publication: model_author · execution: Moonshine AI · checked 2026-10-19. Author-run eight-set evaluation. Exact cleaned-reference versions/checkpoint commits are unspecified. Table 2 is end-of-utterance response latency on an Apple MacBook M3, not RTFx or time to first token. Parameter counts differ from current Hub tensor metadata and are retained separately.
moonshine-ai/moonshine-streaming-tiny model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: UsefulSensors/moonshine-streaming-tiny. Model commit and evaluation date are unspecified unless the observation records them.
moonshine-ai/moonshine-streaming-tiny Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
moonshine-ai/moonshine-streaming-small model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: UsefulSensors/moonshine-streaming-small. Model commit and evaluation date are unspecified unless the observation records them.
moonshine-ai/moonshine-streaming-small Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
moonshine-ai/moonshine-streaming-medium model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: UsefulSensors/moonshine-streaming-medium. Model commit and evaluation date are unspecified unless the observation records them.
moonshine-ai/moonshine-streaming-medium Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
moonshine-ai/moonshine-tiny model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: UsefulSensors/moonshine-tiny. Model commit and evaluation date are unspecified unless the observation records them.
moonshine-ai/moonshine-tiny Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
moonshine-ai/moonshine-base model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: UsefulSensors/moonshine-base. Model commit and evaluation date are unspecified unless the observation records them.
moonshine-ai/moonshine-base Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
openai/whisper-tiny model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: openai/whisper-tiny. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-tiny Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
openai/whisper-base model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: openai/whisper-base. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-base Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
openai/whisper-small model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: openai/whisper-small. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-small Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
openai/whisper-medium model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: openai/whisper-medium. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-medium Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
openai/whisper-tiny.en model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: openai/whisper-tiny.en. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-tiny.en Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
openai/whisper-base.en model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: openai/whisper-base.en. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-base.en Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
openai/whisper-small.en model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: openai/whisper-small.en. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-small.en Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
openai/whisper-large-v2 model card
Publication: model_author · execution: not_applicable · checked 2026-10-19. Exact benchmark row: openai/whisper-large-v2. Model commit and evaluation date are unspecified unless the observation records them.
openai/whisper-large-v2 Hub tensor metadata
Publication: platform_metadata · execution: not_applicable · checked 2026-10-19. Hub safetensors parameter count, not a download-size or RAM estimate. May count stored/tied tensors differently from the model card.
transcribe.cpp publication benchmark profile
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Q8_0 FLEURS test and LibriSpeech test-clean, with per-cell quantization, batch size, engine revision, uncertainty and sample counts. Legacy-published rows may omit run dates and engine SHAs. Single-clip speed checks are ancillary and never create aggregate votes.
transcribe.cpp breeze-asr-25 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp canary-180m-flash Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp canary-1b-flash Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp canary-1b-v2 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp canary-1b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp canary-qwen-2.5b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp cohere-transcribe-03-2026 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp cohere-transcribe-arabic-07-2026 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp fun-asr-mlt-nano-2512 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp fun-asr-nano-2512 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp granite-4.0-1b-speech Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp granite-speech-4.1-2b-nar Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp granite-speech-4.1-2b-plus Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp granite-speech-4.1-2b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp Granite Speech 5.0 470M TurboCTC NC Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp Granite Speech 5.0 470M TurboCTC Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp medasr Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp moonshine-base Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp moonshine-streaming-medium Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp moonshine-streaming-small Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp moonshine-streaming-tiny Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp moonshine-tiny Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp moss-transcribe-diarize Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp multitalker-parakeet-streaming-0.6b-v1 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp nemotron-3.5-asr-streaming-0.6b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp nemotron-speech-streaming-en-0.6b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp Orukeet Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-ctc-0.6b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-ctc-1.1b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-primeline Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-rnnt-0.6b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-rnnt-1.1b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-tdt-0.6b-v2 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-tdt-0.6b-v3 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-tdt-1.1b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-tdt_ctc-1.1b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp parakeet-tdt_ctc-110m Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp Parakeet Ultra Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp Parakeet Unified EN 0.6B Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp Phonon-2 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp qwen3-asr-0.6b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp qwen3-asr-1.7b Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp sensevoice-small Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp voxtral-mini-3b-2507 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp voxtral-mini-4b-realtime-2602 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp voxtral-small-24b-2507 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-base.en Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-base Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-large-v2 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp Whisper large-v3-turbo Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-large-v3 Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-large Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-medium.en Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-medium Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-small.en Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-small Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-tiny.en Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
transcribe.cpp whisper-tiny Q8_0 catalog
Publication: runtime_benchmark_maintainer · execution: unspecified · checked 2026-10-19. Published runtime benchmark, not the upstream full-precision checkpoint. Original measurement provenance, engine SHA, CI95, batch size and utterance counts are retained. Publication does not identify a particular execution operator.
Deepgram official speech API pricing
Publication: vendor · execution: not_applicable · checked 2026-10-19. Pay-as-you-go standard rates are separate from Growth discounts and current streaming promotions; optional diarization/keyterm add-ons are excluded. Minute rates are normalized only by multiplying by 60.
AssemblyAI official API pricing
Publication: vendor · execution: not_applicable · checked 2026-10-19. Pre-recorded Pay As You Go Universal-3.5 Pro is $0.21 per audio hour; Universal-2 is $0.15. Optional add-ons and realtime products are separate.
Gladia official speech API pricing
Publication: vendor · execution: not_applicable · checked 2026-10-19. Self-serve async Solaria-3, 0–100 hours/month, is EUR 0.16 per hour. Live service is EUR 0.68 per hour. Currency stays EUR; no USD exchange-rate assumption.
Mistral Voxtral Mini Transcribe 2602 official model pricing
Publication: vendor · execution: not_applicable · checked 2026-10-19. USD 0.003 per minute for hosted Voxtral Mini Transcribe 2602. This is not the open Voxtral Mini 4B Realtime 2602 checkpoint.
Soniox official token pricing
Publication: vendor · execution: not_applicable · checked 2026-10-19. The page gives approximate hourly equivalents, not flat hourly tariffs. Actual per-million audio/text token rates are preserved; no hourly price is inferred.
Open ASR multilingual prepared test datasets
Publication: benchmark_maintainer · execution: not_applicable · checked 2026-10-19. Prepared FLEURS, MCV and MLS test configurations use corpus_language names. Exact dataset commit for each published run is unspecified. Armenian is hosted separately; Hindi uses public Monsoon lattice references and voi-oiwer, as documented in Open ASR constants.