Best AI for transcription
Word error rate is lower-is-better and inverted before ranking. RTFx is carried as a separate speed benchmark, never folded into accuracy.
Rumeqo runs these models inside your team rooms. See what each one costs.
91 of 91 ranked models
| rank | model | vendor | composite | benchmarks | % wer, lower is better | % wer, lower is better | % wer, lower is better | % wer, lower is better | % wer, lower is better | % wer, lower is better | % wer, lower is better | rtfx |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Scribe v2elevenlabs/scribe-v2 | ElevenLabs | 86.4 | 7 of 8 benchmarks | 4.7% | 7.3% | 3.3% | 2.3% | 2.3% | 2.5% | 2.9% | |
| 2 | Azure Speechmicrosoft/azure-speech | Microsoft | 85.8 | 6 of 8 benchmarks | 4.5% | 3.3% | 1.9% | 2.3% | 2.5% | 3.1% | ||
| 3 | Resonant 1reson8/resonant-1 | Reson8 | 79.7 | 7 of 8 benchmarks | 4.6% | 8.5% | 4.1% | 2.8% | 2.7% | 3.0% | 3.5% | |
| 4 | Resonant 1 Flashreson8/resonant-1-flash | Reson8 | 79.4 | 7 of 8 benchmarks | 4.7% | 8.4% | 4.1% | 2.8% | 2.7% | 3.0% | 3.5% | |
| 5 | Universal 3 Proassemblyai/universal-3-pro | AssemblyAI | 78.0 | 7 of 8 benchmarks | 5.2% | 8.3% | 3.8% | 2.6% | 2.5% | 4.1% | 3.9% | |
| 6 | Cohere Transcribecohere/cohere-transcribe | Cohere | 68.4 | 8 of 8 benchmarks | 5.2% | 9.7% | 4.0% | 3.1% | 2.8% | 3.0% | 6.0% | 915.6x |
| 7 | Multilingualmodulate/multilingual | Modulate | 67.4 | 5 of 8 benchmarks | 4.1% | 3.5% | 3.0% | 3.2% | 4.9% | |||
| 8 | Vfastmodulate/vfast | Modulate | 65.5 | 1 of 8 benchmarks | 4.4% | |||||||
| 9 | Parakeet Tdt 0.6B v2nvidia/parakeet-tdt-0.6b-v2 | NVIDIA | 63.5 | 3 of 8 benchmarks | 5.4% | 11.2% | 6038.1x | |||||
| 10 | Enhancedspeechmatics/enhanced | Speechmatics | 62.5 | 7 of 8 benchmarks | 5.9% | 8.8% | 4.9% | 2.8% | 2.7% | 5.4% | 5.0% | |
| 11 | Scribe v1zoom/scribe-v1 | Zoom | 62.5 | 1 of 8 benchmarks | 4.7% | |||||||
| 12 | Voxtral Small 24B 2507mistralai/voxtral-small-24b-2507 | Mistral | 61.1 | 7 of 8 benchmarks | 5.7% | 4.1% | 2.9% | 2.9% | 3.8% | 4.4% | 100.1x | |
| 13 | STT Async v5soniox/stt-async-v5 | Soniox | 60.4 | 5 of 8 benchmarks | 6.2% | 3.6% | 3.4% | 4.2% | 4.0% | |||
| 14 | Fusionrev/fusion | Rev | 58.8 | 1 of 8 benchmarks | 9.5% | |||||||
| 15 | Parakeet TDT 0.6B v3nvidia/parakeet-tdt-0.6b-v3 | NVIDIA | 58.7 | 8 of 8 benchmarks | 5.7% | 10.7% | 5.4% | 4.1% | 3.7% | 4.7% | 6.1% | 6098.2x |
| 16 | Universal 3.5 Proassemblyai/universal-3-5-pro | AssemblyAI | 58.6 | 1 of 8 benchmarks | 5.0% | |||||||
| 17 | Canary Qwen 2.5Bnvidia/canary-qwen-2.5b | NVIDIA | 57.9 | 3 of 8 benchmarks | 5.1% | 11.2% | 861.5x | |||||
| 18 | Machinerev/machine | Rev | 57.5 | 1 of 8 benchmarks | 9.6% | |||||||
| 19 | ARK ASR 3Bautoark-ai/ark-asr-3b | AutoArk AI | 56.4 | 2 of 8 benchmarks | 4.6% | 484.3x | ||||||
| 20 | Parakeet Tdt CTC (110M)nvidia/parakeet-tdt-ctc:110m | NVIDIA | 55.6 | 2 of 8 benchmarks | 6.6% | 6118.8x | ||||||
| 21 | Granite Speech 4.1 2Bibm-granite/granite-speech-4.1-2b | IBM | 55.6 | 2 of 8 benchmarks | 4.9% | 546.8x | ||||||
| 22 | Granite Speech 4.1 2B NARibm-granite/granite-speech-4.1-2b-nar | IBM | 55.4 | 6 of 8 benchmarks | 5.0% | 5.4% | 4.3% | 3.6% | 6.5% | 2079.3x | ||
| 23 | Moonshine Streaming (medium)usefulsensors/moonshine-streaming:medium | Useful Sensors | 55.3 | 2 of 8 benchmarks | 5.8% | 2681.0x | ||||||
| 24 | Avalon v1 ENaquavoice/avalon-v1-en | Aqua Voice | 54.9 | 1 of 8 benchmarks | 5.2% | |||||||
| 25 | Canary 1B Flashnvidia/canary-1b-flash | NVIDIA | 54.1 | 2 of 8 benchmarks | 5.8% | 2126.1x | ||||||
| 26 | Canary 1B v2nvidia/canary-1b-v2 | NVIDIA | 54.0 | 7 of 8 benchmarks | 6.4% | 4.8% | 4.1% | 3.2% | 4.8% | 6.3% | 1821.4x | |
| 27 | Solaria 3upstage/solaria-3 | Upstage | 54.0 | 1 of 8 benchmarks | 5.3% | |||||||
| 28 | ARK ASR 0.6Bautoark-ai/ark-asr-0.6b | AutoArk AI | 54.0 | 2 of 8 benchmarks | 5.1% | 672.2x | ||||||
| 29 | Granite 4.0 1B Speechibm-granite/granite-4.0-1b-speech | IBM | 53.9 | 2 of 8 benchmarks | 5.1% | 661.2x | ||||||
| 30 | Nemotron Speech Streaming EN 0.6Bnvidia/nemotron-speech-streaming-en-0.6b | NVIDIA | 53.4 | 2 of 8 benchmarks | 5.7% | 1071.4x | ||||||
| 31 | Pulsesmallestai/pulse | Smallest AI | 52.9 | 1 of 8 benchmarks | 5.4% | |||||||
| 32 | Parakeet CTC 1.1Bnvidia/parakeet-ctc-1.1b | NVIDIA | 52.2 | 3 of 8 benchmarks | 6.5% | 12.9% | 5014.5x | |||||
| 33 | Qwen3 ASR 1.7Bqwen/qwen3-asr-1.7b | Qwen | 52.1 | 2 of 8 benchmarks | 5.0% | 394.1x | ||||||
| 34 | Phi 4 Multimodal Instructmicrosoft/phi-4-multimodal-instruct | Microsoft | 52.1 | 7 of 8 benchmarks | 5.4% | 5.0% | 4.1% | 3.8% | 4.3% | 5.2% | 162.7x | |
| 35 | Canary 180M Flashnvidia/canary-180m-flash | NVIDIA | 51.3 | 2 of 8 benchmarks | 6.3% | 2484.2x | ||||||
| 36 | Qwen3 ASR 1.7B HFqwen/qwen3-asr-1.7b-hf | Qwen | 51.0 | 7 of 8 benchmarks | 5.0% | 5.7% | 4.0% | 3.8% | 5.4% | 6.3% | 796.2x | |
| 37 | Whisper Large V3openai/whisper-large-v3 | OpenAI | 50.9 | 8 of 8 benchmarks | 6.5% | 11.2% | 6.2% | 4.0% | 3.4% | 4.5% | 4.9% | 462.2x |
| 38 | MOSS Transcribe Preview 2Bopenmoss/moss-transcribe-preview-2b | OpenMOSS | 50.3 | 2 of 8 benchmarks | 4.7% | 148.5x | ||||||
| 39 | MOSS Transcribe Diarizeopenmoss/moss-transcribe-diarize | OpenMOSS | 50.0 | 2 of 8 benchmarks | 5.2% | 382.4x | ||||||
| 40 | Higgs Audio v3 STTbosonai/higgs-audio-v3-stt | Boson AI | 49.9 | 2 of 8 benchmarks | 4.6% | 110.1x | ||||||
| 41 | Parakeet CTC 0.6Bnvidia/parakeet-ctc-0.6b | NVIDIA | 49.9 | 3 of 8 benchmarks | 6.7% | 13.7% | 5883.9x | |||||
| 42 | Hojo ASR v1hojoai/hojo-asr-v1 | Hojo AI | 49.6 | 2 of 8 benchmarks | 4.5% | 73.8x | ||||||
| 43 | Moonshine Streaming Smallusefulsensors/moonshine-streaming-small | Useful Sensors | 49.3 | 2 of 8 benchmarks | 6.8% | 3206.0x | ||||||
| 44 | Distil Large v3.5distil-whisper/distil-large-v3.5 | Distil Whisper | 49.2 | 3 of 8 benchmarks | 6.1% | 11.7% | 874.0x | |||||
| 45 | Parakeet Tdt 1.1Bnvidia/parakeet-tdt-1.1b | NVIDIA | 49.0 | 3 of 8 benchmarks | 6.2% | 15.8% | 4528.7x | |||||
| 46 | Canary 1Bnvidia/canary-1b | NVIDIA | 48.9 | 2 of 8 benchmarks | 5.8% | 765.6x | ||||||
| 47 | Granite Speech 3.3 2Bibm-granite/granite-speech-3.3-2b | IBM | 48.6 | 2 of 8 benchmarks | 5.5% | 507.1x | ||||||
| 48 | Higgs Audio v3 8B STT v2bosonai/higgs-audio-v3-8b-stt-v2 | Boson AI | 48.5 | 2 of 8 benchmarks | 4.7% | 139.1x | ||||||
| 49 | Niagara 38M Batch.enabr-ai/niagara-38m-batch.en | Applied Brain Research | 47.8 | 2 of 8 benchmarks | 8.3% | 4048.7x | ||||||
| 50 | OmniASR LLM 7B v2meta-llama/omniasr-llm-7b-v2 | Meta | 47.7 | 7 of 8 benchmarks | 6.9% | 5.3% | 4.4% | 3.4% | 3.6% | 5.2% | 142.9x | |
| 51 | Parakeet Rnnt 0.6Bnvidia/parakeet-rnnt-0.6b | NVIDIA | 47.7 | 3 of 8 benchmarks | 6.7% | 14.5% | 5406.7x | |||||
| 52 | Moonshine Streaming Tinyusefulsensors/moonshine-streaming-tiny | Useful Sensors | 47.6 | 2 of 8 benchmarks | 11.2% | 4375.2x | ||||||
| 53 | Granite Speech 3.3 8Bibm-granite/granite-speech-3.3-8b | IBM | 47.4 | 2 of 8 benchmarks | 5.3% | 263.7x | ||||||
| 54 | Niagara 19M Batch.enabr-ai/niagara-19m-batch.en | Applied Brain Research | 46.7 | 2 of 8 benchmarks | 9.9% | 3735.5x | ||||||
| 55 | Qwen3 ASR 0.6Bqwen/qwen3-asr-0.6b | Qwen | 46.7 | 2 of 8 benchmarks | 5.6% | 438.7x | ||||||
| 56 | Parakeet Rnnt 1.1Bnvidia/parakeet-rnnt-1.1b | NVIDIA | 46.4 | 3 of 8 benchmarks | 6.4% | 17.1% | 4121.8x | |||||
| 57 | Zipformer Cr CTC Transducer XL (290M)soundsgoodai/zipformer-cr-ctc-transducer-xl:290m | SoundsGood AI | 46.1 | 2 of 8 benchmarks | 5.2% | 159.8x | ||||||
| 58 | Moonshine Tinyusefulsensors/moonshine-tiny | Useful Sensors | 45.6 | 2 of 8 benchmarks | 11.4% | 3733.1x | ||||||
| 59 | Chirpgoogle/chirp | 45.5 | 1 of 8 benchmarks | 13.1% | ||||||||
| 60 | Whisper Large V3 Turboopenai/whisper-large-v3-turbo | OpenAI | 44.9 | 8 of 8 benchmarks | 7.0% | 11.0% | 6.7% | 6.1% | 3.9% | 5.4% | 4.6% | 782.6x |
| 61 | Hubert Large Ls960 Ftmeta-llama/hubert-large-ls960-ft | Meta | 44.1 | 2 of 8 benchmarks | 13.4% | 3024.0x | ||||||
| 62 | Wav2vec2 Large 960h Lv60 Selfmeta-llama/wav2vec2-large-960h-lv60-self | Meta | 44.0 | 2 of 8 benchmarks | 11.8% | 3010.0x | ||||||
| 63 | STT EN Conformer Transducer Smallnvidia/stt-en-conformer-transducer-small | NVIDIA | 42.8 | 1 of 8 benchmarks | 13.7% | |||||||
| 64 | Owsm CTC v4 1Bespnet/owsm-ctc-v4-1b | ESPnet | 42.6 | 2 of 8 benchmarks | 6.7% | 764.7x | ||||||
| 65 | Voxtral Mini 3B 2507mistralai/voxtral-mini-3b-2507 | Mistral AI | 42.5 | 7 of 8 benchmarks | 6.0% | 6.0% | 4.5% | 4.2% | 5.1% | 5.1% | 179.7x | |
| 66 | Owsm CTC v3.1 1Bespnet/owsm-ctc-v3.1-1b | ESPnet | 42.0 | 2 of 8 benchmarks | 7.3% | 816.2x | ||||||
| 67 | GLM ASR Nano 2512z-ai/glm-asr-nano-2512 | Z.ai | 41.8 | 2 of 8 benchmarks | 6.0% | 333.0x | ||||||
| 68 | STT EN Conformer CTC Largenvidia/stt-en-conformer-ctc-large | NVIDIA | 41.4 | 1 of 8 benchmarks | 14.4% | |||||||
| 69 | Mms 1B Allmeta-llama/mms-1b-all | Meta | 41.3 | 2 of 8 benchmarks | 13.5% | 1954.0x | ||||||
| 70 | Audio8 ASR 0.1Bautoark-ai/audio8-asr-0.1b | AutoArk AI | 40.1 | 2 of 8 benchmarks | 7.0% | 709.3x | ||||||
| 71 | STT 2.6B ENkyutai/stt-2.6b-en | Kyutai | 39.6 | 2 of 8 benchmarks | 5.7% | 133.2x | ||||||
| 72 | Owsm CTC v3.2 Ft 1Bespnet/owsm-ctc-v3.2-ft-1b | ESPnet | 39.2 | 2 of 8 benchmarks | 7.3% | 692.3x | ||||||
| 73 | OmniASR LLM 3B v2meta-llama/omniasr-llm-3b-v2 | Meta | 39.1 | 5 of 8 benchmarks | 6.9% | 5.6% | 4.0% | 5.0% | 7.1% | |||
| 74 | Lite Whisper Large v3 Accefficient-speech/lite-whisper-large-v3-acc | Efficient Speech | 39.1 | 2 of 8 benchmarks | 6.3% | 203.9x | ||||||
| 75 | Zipformer Transducer XL (290M)soundsgoodai/zipformer-transducer-xl:290m | SoundsGood AI | 37.3 | 2 of 8 benchmarks | 6.2% | 141.2x | ||||||
| 76 | Crisperwhispernyrahealth/crisperwhisper | Nyra Health | 36.8 | 2 of 8 benchmarks | 5.8% | 33.3x | ||||||
| 77 | STT EN Conformer CTC Smallnvidia/stt-en-conformer-ctc-small | NVIDIA | 36.1 | 1 of 8 benchmarks | 18.7% | |||||||
| 78 | OmniASR LLM 1B v2meta-llama/omniasr-llm-1b-v2 | Meta | 35.2 | 5 of 8 benchmarks | 7.0% | 6.3% | 4.3% | 4.8% | 7.3% | |||
| 79 | OmniASR CTC 7B v2meta-llama/omniasr-ctc-7b-v2 | Meta | 35.0 | 7 of 8 benchmarks | 9.0% | 7.6% | 6.0% | 4.4% | 4.7% | 6.5% | 527.6x | |
| 80 | STT EN Fastconformer CTC Largenvidia/stt-en-fastconformer-ctc-large | NVIDIA | 34.8 | 1 of 8 benchmarks | 21.5% | |||||||
| 81 | STT EN Conformer Transducer Largenvidia/stt-en-conformer-transducer-large | NVIDIA | 33.5 | 1 of 8 benchmarks | 21.9% | |||||||
| 82 | OmniASR CTC 3B v2meta-llama/omniasr-ctc-3b-v2 | Meta | 33.2 | 5 of 8 benchmarks | 8.1% | 6.2% | 4.5% | 5.1% | 6.5% | |||
| 83 | STT EN Fastconformer Transducer Largenvidia/stt-en-fastconformer-transducer-large | NVIDIA | 32.1 | 1 of 8 benchmarks | 22.5% | |||||||
| 84 | Voxtral Mini 4B Realtime 2602mistralai/voxtral-mini-4b-realtime-2602 | Mistral AI | 30.4 | 7 of 8 benchmarks | 6.4% | 7.8% | 6.3% | 4.2% | 6.4% | 6.2% | 105.1x | |
| 85 | Qwen3 ASR 0.6B HFqwen/qwen3-asr-0.6b-hf | Qwen | 28.7 | 7 of 8 benchmarks | 5.6% | 8.8% | 6.7% | 5.8% | 9.1% | 10.1% | 730.2x | |
| 86 | ASR Conformer Largescaleasrspeechbrain/asr-conformer-largescaleasr | SpeechBrain | 27.6 | 2 of 8 benchmarks | 7.6% | 72.4x | ||||||
| 87 | Nemotron 3.5 ASR Streaming 0.6Bnvidia/nemotron-3.5-asr-streaming-0.6b | NVIDIA | 26.5 | 7 of 8 benchmarks | 7.9% | 9.4% | 9.0% | 5.4% | 8.7% | 8.3% | 1489.6x | |
| 88 | OmniASR CTC 1B v2meta-llama/omniasr-ctc-1b-v2 | Meta | 23.0 | 5 of 8 benchmarks | 10.3% | 8.4% | 5.6% | 6.4% | 9.0% | |||
| 89 | OmniASR LLM 300M v2meta-llama/omniasr-llm-300m-v2 | Meta | 20.5 | 5 of 8 benchmarks | 10.0% | 9.1% | 5.8% | 6.9% | 9.5% | |||
| 90 | Vibevoice ASR HFmicrosoft/vibevoice-asr-hf | Microsoft | 20.4 | 7 of 8 benchmarks | 6.3% | 14.0% | 14.2% | 6.7% | 12.3% | 7.8% | 221.2x | |
| 91 | OmniASR CTC 300M v2meta-llama/omniasr-ctc-300m-v2 | Meta | 14.3 | 5 of 8 benchmarks | 18.6% | 16.5% | 9.9% | 11.5% | 14.4% |
How this ranks
Every benchmark value becomes a percentile among the models that have it, so accuracy scores, Elo ratings and word error rates compare without hand-tuned scaling. Metrics where lower is better are inverted first. Raw values are never summed or averaged across benchmarks. A model's mean percentile is then shrunk toward the mean of the models that were broadly benchmarked, so a model tested twice cannot outrank a broadly tested one on two lucky results. Turning a data source off runs that same ranking code again in your browser over the sources you left on.
A model scored on fewer than 2 of the 8 ranked benchmarks in this category still ranks here, on the benchmarks it does have, and its row carries a partial coverage mark. On an equal score it sits under the model that earned the same number across more of the board.
Data sources
Turn a source off to drop every benchmark it feeds and rank the board again from what is left, in your browser. Turn them all off and the table has nothing to rank. Your choice follows you across the leaderboard pages.
- Open ASR Leaderboard
Word error rates from the Hugging Face Open ASR Leaderboard.