Fermion ResearchFermion Research

Speech recognition/facts checked 13 November 2026

The best open-source speech-to-text model for your use

Open-weight speech recognition models now come within a fraction of a point of each other on English. Size, language coverage and where the model runs decide more than the leaderboard average does.

Phonon-2.5 downloads and benchmarks

The short answer

There is no single best open-source speech-to-text model: on the seven English test sets of the Open ASR Leaderboard (snapshot 2026-09-25), the lowest open-weight averages are Hojo-ASR-V1 at 4.29 % word error and Qwen3-ASR 1.7B at 4.43 %, while Whisper large-v3 still covers the most languages, 99. The right choice depends on whether you need the lowest English error, the smallest download, or many languages.

For small downloads, Phonon-2.5 English (158 MB) averages 4.82 % on the same seven sets in our run with the leaderboard's code, the same average as the 2,472 MB Parakeet TDT 0.6B v2, and the 25-language Phonon-2.5 (156 MB) averages 5.06 %. The lowest averages on the whole board belong to closed APIs, for example Zoom Scribe v2 Pro at 3.68 %.

Open-weight models compared

Word error rate in %, lower is better, averaged over the seven classic Open ASR Leaderboard English sets. Sizes are as listed in the snapshot table. Language counts are from each model's own card.

Open ASR Leaderboard seven-set English average (snapshot 2026-09-25); Phonon rows are Fermion Research runs with the leaderboard's code
ModelSizeLanguages (model card)English 7-set WER
Hojo-ASR-V110,360 MBMandarin, English, Cantonese, Sichuan dialect4.29
Qwen3-ASR 1.7B4,080 MB30 languages and 22 Chinese dialects4.43
Higgs Audio v3 STT5,360 MBnot stated on the card4.45
Canary-Qwen 2.5B5,000 MBEnglish4.55
ARK-ASR-3B7,500 MB194.62
Granite Speech 4.1 2B4,000 MB64.69
MOSS-Transcribe-Diarize1,817 MB50+, with speaker labels4.75
Parakeet TDT 0.6B v22,472 MBEnglish4.82
Phonon-2.5 English158 MBEnglish4.82
Cohere Transcribe 03-20264,000 MB144.85
Parakeet TDT 0.6B v32,508 MB25 European4.96
Phonon-2.5156 MB25 European5.06
Granite Speech 5.0 470M TurboCTC946 MB75.08
Voxtral Mini 3B10,000 MB8 named on the card5.63
Canary 1B v23,916 MB255.81
Whisper large-v33,090 MB996.04
Whisper large-v3-turbo1,618 MB996.58
Niagara 84M168 MBEnglish7.27

Which model is best for what

  • Lowest English error with a GPU: Hojo-ASR-V1 and Qwen3-ASR 1.7B lead the open rows. Hojo-ASR-V1 covers Mandarin, English and Cantonese; Qwen3-ASR covers 30 languages.
  • Most languages: Whisper large-v3 and large-v3-turbo (99 languages). MOSS-Transcribe-Diarize states 50+ languages and labels speakers in the same pass.
  • European languages from a small download: Phonon-2.5, 25 languages in 156 MB. Parakeet TDT 0.6B v3 and Canary 1B v2 cover 25 languages at 2.5 to 3.9 GB.
  • English on a laptop, phone or small container: Phonon-2.5 English (158 MB, 4.82 %) or Niagara 84M (168 MB in the snapshot, 7.27 %, English only, built for CPUs according to its card).
  • Speaker labels: MOSS-Transcribe-Diarize. The Phonon models do not label speakers.
  • Translation: Whisper (except turbo) and Canary 1B v2 translate speech; Phonon transcribes only.

How to read a leaderboard average

An average over seven test sets hides a lot. AMI is meeting audio recorded at a distance, Earnings-22 is financial calls with numbers and names, LibriSpeech is read audiobooks. Two models with the same average can differ by several points on the set that looks like your audio, so check the per-set columns on the leaderboard before choosing.

Differences of a tenth of a point are within the noise between runs and normalisers. Treat models within about 0.2 points as tied and decide on size, languages and runtime.

The leaderboard measures English only. For other languages, FLEURS is the common public test; our multilingual guide has per-language numbers for three 25-language models.

Questions

What is the most accurate open-source speech-to-text model?
On the seven English Open ASR Leaderboard sets (snapshot 2026-09-25), Hojo-ASR-V1 has the lowest open-weight average at 4.29 % word error, followed by Qwen3-ASR 1.7B at 4.43 %. Closed APIs such as Zoom Scribe v2 Pro (3.68 %) are lower still.
What is the best small open speech-to-text model?
Phonon-2.5 English, 158 MB, averages 4.82 % on the seven English sets in our run with the leaderboard's code, the same as the 2,472 MB Parakeet TDT 0.6B v2. For 25 languages, Phonon-2.5 is 156 MB and averages 5.06 %.
Is Whisper still the best open speech model?
Whisper large-v3 covers 99 languages and translates speech to English, which no smaller model in this table matches. On English accuracy it averages 6.04 % on the seven sets, behind most newer open models.
Which open model is best for European languages?
Phonon-2.5, Parakeet TDT 0.6B v3 and Canary 1B v2 each cover 25 European languages. On FLEURS, Phonon-2.5 averages 11.71 % across its 25 languages and Parakeet TDT 0.6B v3 13.26 %, measured with one scoring code.
Do these numbers apply to my audio?
Only roughly. Public test sets are a common reference, but accents, microphones and vocabulary change the error rate, so run two or three candidates on a sample of your own recordings.
Which open model labels speakers?
Among the models in this table, MOSS-Transcribe-Diarize performs transcription and speaker diarization in one model, according to its card. The Phonon models transcribe without speaker labels.

Related

Sources and method

  • Rival English averages: Open ASR Leaderboard, seven English sets (LibriSpeech clean and other, AMI, Earnings-22, GigaSpeech, SPGISpeech, VoxPopuli), snapshot 2026-09-25.
  • Phonon English averages: Fermion Research measurement with the Open ASR Leaderboard's scoring code on the full seven test sets (Phonon-2.5 release post, 2026-10-22).
  • Language coverage: the Hugging Face model cards of HojoAI/Hojo-ASR-V1, Qwen/Qwen3-ASR-1.7B, bosonai/higgs-audio-v3-stt, nvidia/canary-qwen-2.5b, AutoArk-AI/ARK-ASR-3B, ibm-granite/granite-speech-4.1-2b, OpenMOSS-Team/MOSS-Transcribe-Diarize, nvidia/parakeet-tdt-0.6b-v2, CohereLabs/cohere-transcribe-03-2026, nvidia/parakeet-tdt-0.6b-v3, ibm-granite/granite-speech-5.0-470m-turboctc, mistralai/Voxtral-Mini-3B-2507, nvidia/canary-1b-v2, openai/whisper-large-v3, openai/whisper-large-v3-turbo, abr-ai/niagara-84m-batch.en; all read 2026-11-13.
  • FLEURS numbers: Fermion Research measurement on the full FLEURS test set of each language, one scoring code and normalisation for every model (Phonon-2.5 release post, 2026-10-22).

Canonical: https://www.fermionresearch.com/speech/best-open-source-speech-to-text-model/