Fermion ResearchFermion Research

Phonon/facts checked 13 November 2026

Phonon-2.5 vs Meta Muse Voice Transcribe

On the leaderboard's English sets the three are close: Muse is just ahead of the 25-language Phonon-2.5 and just behind Phonon-2.5 English. The bigger differences are hosting and features.

Phonon-2.5 downloads and benchmarks

The short answer

On the Open ASR Leaderboard's seven English sets, Meta Muse Voice Transcribe averages 4.93 % word error, a little lower than Phonon-2.5 (5.06 %) and a little higher than Phonon-2.5 English (4.82 %). Choose Muse if you want a hosted streaming API with speaker attribution; choose Phonon when you need the model on your own hardware, offline, or without an hourly fee.

The set-by-set picture splits cleanly: Phonon is lower on LibriSpeech and AMI meetings, Muse is lower on earnings calls, GigaSpeech, SPGISpeech and VoxPopuli.

Side by side

Meta facts from Meta's model page (dev.meta.ai/models/muse-voice-transcribe), read 2026-11-13.

Phonon-2.5Meta Muse Voice Transcribe
Model typeOpen weights on Hugging FaceHosted API (Meta Model API), model ID muse-voice-transcribe-1.0
Size156 MB download (Phonon-2.5 English: 158 MB)Not published
Languages25 European languages (Phonon-2.5 English: English only)Not listed on Meta's model page
English word error, seven Open ASR sets5.06 % (Phonon-2.5 English: 4.82 %)4.93 %
Price per audio hourNo per-hour fee when self-hosted$0.18 ($3.00 per 1,000 minutes)
Where it runsYour hardware: CPUs (Linux x86-64 and arm64, Windows x86-64), Apple silicon, NVIDIA GPUs, AndroidMeta's cloud
StreamingYes, live WebSocket from fermion serveYes, with end-of-speech detection in the model
DiarizationNoSpeaker attribution for 20+ speakers
Word timestampsYes (--json output)Not stated on Meta's model page

English, set by set

Phonon-2.5 and Phonon-2.5 English are lower on LibriSpeech clean, LibriSpeech other and AMI. Muse is lower on Earnings-22, 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 EnglishMeta Muse Voice Transcribe
LibriSpeech clean1.601.641.84
LibriSpeech other3.413.564.51
AMI9.128.769.20
Earnings-226.726.226.01
GigaSpeech8.047.997.75
SPGISpeech3.313.313.21
VoxPopuli3.232.282.02
Average of the seven5.064.824.93

Where Muse Voice Transcribe is the better choice

  • Live conversations with several people: speaker attribution for 20 or more speakers and end-of-speech detection inside the recognition model.
  • Contextual and keyword biasing for custom vocabulary.
  • Earnings calls, podcasts and parliamentary audio, where its per-set scores are lower.
  • An API that Meta says works with OpenAI-SDK-compatible clients.

Where Phonon-2.5 is the better choice

  • English accuracy with Phonon-2.5 English: a lower seven-set average (4.82 % against 4.93 %).
  • Read speech and meetings: lower word error on LibriSpeech and AMI.
  • Your hardware, your network: a 156 MB model that runs offline on CPUs, Macs, NVIDIA GPUs and Android.
  • No hourly fee, and the same OpenAI-shaped route (POST /v1/audio/transcriptions) from fermion serve.
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 Meta Muse Voice Transcribe more accurate than Phonon?
It depends on which Phonon. On the Open ASR Leaderboard's seven English sets Muse averages 4.93 %, Phonon-2.5 5.06 % and Phonon-2.5 English 4.82 %; Phonon is lower on LibriSpeech and AMI, Muse on the other four sets.
How much does Muse Voice Transcribe cost?
Meta lists muse-voice-transcribe-1.0 at $3.00 per 1,000 minutes, which is $0.18 per audio hour (read 2026-11-13).
Can I download Muse Voice Transcribe?
Meta offers it through the Meta Model API. Phonon-2.5 and Phonon-2.5 English are open weights on Hugging Face.
Does Phonon-2.5 identify speakers like Muse?
No. Phonon-2.5 returns text with word timestamps and no speaker labels. Muse attributes speech to 20 or more speakers.
Do both stream?
Yes. Muse is a streaming model; Phonon streams through the GET /v1/audio/stream WebSocket of fermion serve.

Related

Sources and method

  • Muse model ID, price, speaker attribution, streaming and biasing: dev.meta.ai/models/muse-voice-transcribe, read 2026-11-13.
  • Rival English word error: Open ASR Leaderboard, seven English sets, snapshot 2026-09-25 (huggingface.co/spaces/hf-audio/open_asr_leaderboard).
  • 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-meta-muse/