Fermion ResearchFermion Research

Phonon/facts checked 13 November 2026

Phonon-2.5 vs Qwen3-ASR

Qwen3-ASR covers 30 languages and 22 Chinese dialects from 1.6 to 4.1 GB. Phonon-2.5 covers 25 European languages from 156 MB. Which one fits depends on your languages.

Phonon-2.5 downloads and benchmarks

The short answer

Qwen3-ASR 1.7B has the lowest English word error of the three: 4.43 % on the Open ASR Leaderboard's seven sets, against 5.06 % for Phonon-2.5, 4.82 % for Phonon-2.5 English and 5.16 % for Qwen3-ASR 0.6B. Choose Qwen3-ASR for Chinese and its dialects, Asian languages, or top English accuracy with a GPU; choose Phonon-2.5 for the European languages Qwen does not list, or when the model must be small or run without a GPU.

Phonon-2.5 covers nine languages that are not in Qwen3-ASR's list: Ukrainian, Slovak, Bulgarian, Croatian, Estonian, Latvian, Lithuanian, Slovenian and Maltese.

Side by side

Qwen3-ASR facts from the Qwen model cards on Hugging Face, read 2026-11-13. Sizes from the leaderboard snapshot.

Phonon-2.5Qwen3-ASR 1.7BQwen3-ASR 0.6B
Model typeOpen weights on Hugging FaceOpen weightsOpen weights
Size156 MB download (Phonon-2.5 English: 158 MB)4,080 MB1,565 MB
Languages25 European languages (Phonon-2.5 English: English only)30 languages plus 22 Chinese dialects and accentsSame as 1.7B
English word error, seven Open ASR sets5.06 % (Phonon-2.5 English: 4.82 %)4.43 %5.16 %
Hosted price per audio hourNo per-hour fee when self-hosted$0.027 (DeepInfra)$0.012 (DeepInfra)
Where it runsYour hardware: CPUs (Linux x86-64 and arm64, Windows x86-64), Apple silicon, NVIDIA GPUs, AndroidGPU with vLLM or TransformersGPU with vLLM or Transformers
StreamingYes, live WebSocket from fermion serveYes, vLLM backend onlyYes, vLLM backend only
DiarizationNoNoNo
Word timestampsYes (--json output)Through a separate forced aligner (11 languages, up to 5 minutes)Same

English, set by set

Qwen3-ASR 1.7B is lower than Phonon-2.5 on all seven sets; Phonon-2.5 English is lower only on VoxPopuli (2.28 % against 2.85 %). Against Qwen3-ASR 0.6B, Phonon-2.5 is lower on LibriSpeech, AMI and Earnings-22 and higher on GigaSpeech, SPGISpeech and VoxPopuli.

Word error rate in %, lower is better. Phonon rows: our run of the leaderboard's code on the full sets. Rival rows: the Open ASR Leaderboard, snapshot 2026-09-25.
Test setPhonon-2.5Phonon-2.5 EnglishQwen3-ASR 1.7BQwen3-ASR 0.6B
LibriSpeech clean1.601.641.261.70
LibriSpeech other3.413.562.944.01
AMI9.128.768.319.33
Earnings-226.726.225.847.83
GigaSpeech8.047.997.227.62
SPGISpeech3.313.312.572.74
VoxPopuli3.232.282.852.90
Average of the seven5.064.824.435.16

Where Qwen3-ASR is the better choice

  • Chinese, Cantonese and 22 dialects and accents, plus Japanese, Korean, Thai, Vietnamese, Hindi, Arabic, Indonesian and more.
  • English accuracy with the 1.7B model.
  • Built-in language identification.
  • Low hosted prices on third-party hosts: DeepInfra lists $0.027 per hour for 1.7B and $0.012 for 0.6B.

Where Phonon-2.5 is the better choice

  • Ukrainian, Slovak, Bulgarian, Croatian, Estonian, Latvian, Lithuanian, Slovenian and Maltese, none of which Qwen3-ASR lists.
  • Size: 156 MB against 1,565 MB and 4,080 MB.
  • No GPU: Phonon runs on laptop and server CPUs; Qwen3-ASR streaming needs the vLLM backend.
  • Word timestamps from the model itself in every language, with no second model.
pip install fermion-research
fermion transcribe phonon-2.5 interview.wav --json
fermion serve phonon-2.5   # POST /v1/audio/transcriptions, GET /v1/audio/stream

Questions

Is Qwen3-ASR more accurate than Phonon-2.5?
Qwen3-ASR 1.7B is, in English: 4.43 % average word error on the Open ASR Leaderboard's seven sets against 5.06 % for Phonon-2.5 and 4.82 % for Phonon-2.5 English. Qwen3-ASR 0.6B averages 5.16 %, higher than both Phonon models.
Which European languages does Qwen3-ASR miss that Phonon-2.5 covers?
Ukrainian, Slovak, Bulgarian, Croatian, Estonian, Latvian, Lithuanian, Slovenian and Maltese are in Phonon-2.5 and not in Qwen3-ASR's list of 30 languages.
How big are the Qwen3-ASR models?
On the Open ASR Leaderboard snapshot, Qwen3-ASR 1.7B is 4,080 MB and Qwen3-ASR 0.6B is 1,565 MB. Phonon-2.5 is a 156 MB download.
Does Qwen3-ASR give word timestamps?
Through a separate model, Qwen3-ForcedAligner-0.6B, which covers 11 languages and up to 5 minutes of speech per call. Phonon prints word timestamps with --json.
Can either run without a GPU?
Phonon-2.5 runs on CPUs, Apple silicon and Android. Qwen's cards describe serving with vLLM or Transformers, and streaming only with vLLM.

Related

Sources and method

  • Qwen3-ASR languages, streaming, aligner and toolkits: huggingface.co/Qwen/Qwen3-ASR-1.7B and huggingface.co/Qwen/Qwen3-ASR-0.6B, read 2026-11-13.
  • Qwen3-ASR hosted prices: OpenRouter model pages (DeepInfra), read 2026-11-11.
  • Rival English word error: Open ASR Leaderboard, seven English sets, snapshot 2026-09-25 (huggingface.co/spaces/hf-audio/open_asr_leaderboard). Sizes from the same snapshot.
  • Phonon-2.5 and Phonon-2.5 English: Fermion Research measurement, Phonon-2.5 release post (2026-10-22), Open ASR Leaderboard scoring code, full test sets of the same seven sets.

Canonical: https://www.fermionresearch.com/models/phonon-2-5/vs-qwen3-asr/