Fermion ResearchFermion Research

Phonon/facts checked 13 November 2026

Phonon-2.5 vs Mistral Voxtral

Mistral's Voxtral models are audio language models: they transcribe, and the larger ones also summarise and answer questions about audio. Phonon-2.5 is a dedicated transcriber at a fraction of the size.

Phonon-2.5 downloads and benchmarks

The short answer

For transcription alone, Phonon-2.5 has a lower English word error than all three open Voxtral models in the Open ASR Leaderboard snapshot: 5.06 % (Phonon-2.5 English 4.82 %) against 5.21 % for Voxtral Small 24B, 5.63 % for Voxtral Mini 3B and 6.12 % for Voxtral Mini 4B Realtime. Choose Voxtral when you want one model that also understands audio (questions, summaries, function calls), or Mistral's hosted transcription with diarization.

The averages hide a split: Voxtral Small 24B is lower on LibriSpeech, SPGISpeech and VoxPopuli, and Phonon is far lower on AMI meeting audio, which carries the average.

Side by side

Voxtral facts from Mistral's Hugging Face model cards and docs.mistral.ai, read 2026-11-13. Hosted prices from OpenRouter, read 2026-11-11.

Phonon-2.5Voxtral Mini 3B (2507)Voxtral Mini 4B Realtime (2602)
Model typeOpen weights on Hugging FaceOpen weightsOpen weights
Size156 MB download (Phonon-2.5 English: 158 MB)About 9.5 GB of GPU memory in bf16 (Mistral's card)4B parameters; a single GPU with 16 GB or more
Languages25 European languages (Phonon-2.5 English: English only)813
English word error, seven Open ASR sets5.06 % (Phonon-2.5 English: 4.82 %)5.63 %6.12 %
Hosted price per audio hourNo per-hour fee when self-hosted$0.06 (DeepInfra)Hosted as Voxtral Mini Transcribe Realtime
Where it runsYour hardware: CPUs (Linux x86-64 and arm64, Windows x86-64), Apple silicon, NVIDIA GPUs, AndroidGPU with vLLM or TransformersGPU with vLLM; Transformers 5.2+
StreamingYes, live WebSocket from fermion serveNoYes, delay configurable from 80 ms to 2.4 s
DiarizationNoNo (Mistral's hosted batch model has it)No
Word timestampsYes (--json output)Hosted batch model onlyNo

English, set by set

Voxtral Small 24B is the closest on average (5.21 %), lower than Phonon-2.5 on four sets. Phonon's lead comes from AMI, where every Voxtral model is above 13 % and Phonon-2.5 is 9.12 %.

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 EnglishVoxtral Small 24BVoxtral Mini 3BVoxtral Mini 4B Realtime
LibriSpeech clean1.601.641.231.481.62
LibriSpeech other3.413.562.803.634.94
AMI9.128.7613.1913.5713.34
Earnings-226.726.227.077.669.31
GigaSpeech8.047.998.338.708.80
SPGISpeech3.313.311.842.152.23
VoxPopuli3.232.282.002.222.60
Average of the seven5.064.825.215.636.12

Hosted Voxtral

Mistral's current hosted transcription models are Voxtral Mini Transcribe 3, Voxtral Small Transcribe 3 and Voxtral Mini Transcribe Realtime 3. Mistral's transcription docs list speaker diarization, word timestamps, context biasing and recordings up to 3 hours for the batch model, and 13 languages. OpenRouter lists Voxtral Mini Transcribe 3.0 at $0.21 and Voxtral Small Transcribe 3.0 at $0.24 per audio hour. These hosted models have no row in our leaderboard snapshot, so no accuracy figure is given for them.

Where each is the better choice

  • Voxtral: questions and summaries about audio, function calling from voice, long recordings in one call (30 minutes for Mini 3B), lower word error on SPGISpeech and VoxPopuli (and on LibriSpeech for Voxtral Small 24B), and hosted diarization.
  • Phonon-2.5: meetings (AMI), 25 European languages including the Baltic and smaller Slavic languages, no GPU needed, and a 156 MB download against several gigabytes.
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 Phonon-2.5 more accurate than Voxtral?
On the Open ASR Leaderboard's seven English sets, Phonon-2.5's average (5.06 %) is lower than Voxtral Small 24B (5.21 %), Voxtral Mini 3B (5.63 %) and Voxtral Mini 4B Realtime (6.12 %). Voxtral models are lower on SPGISpeech and VoxPopuli (Voxtral Small 24B also on LibriSpeech); Phonon is much lower on AMI.
Can Phonon-2.5 answer questions about audio like Voxtral?
No. Phonon-2.5 is a transcription model. Voxtral Mini 3B also answers questions, summarises and calls functions from audio.
Does Voxtral need a GPU?
Mistral's cards describe GPU serving: about 9.5 GB of GPU memory for Voxtral Mini 3B in bf16 and a GPU with 16 GB or more for Voxtral Mini 4B Realtime. Phonon-2.5 runs on a CPU.
Which languages does Voxtral support?
Mistral's card lists 8 languages for Voxtral Mini 3B and 13 for Voxtral Mini 4B Realtime and the hosted transcription models. Phonon-2.5 covers 25 European languages.
Which is better for live transcription?
Both stream. Voxtral Mini 4B Realtime has a configurable delay from 80 ms to 2.4 s on a GPU; Phonon streams over a WebSocket from fermion serve on CPU or GPU.

Related

Sources and method

  • Voxtral Mini 3B: huggingface.co/mistralai/Voxtral-Mini-3B-2507, read 2026-11-13. Voxtral Mini 4B Realtime: huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602, read 2026-11-13.
  • Hosted models and features: docs.mistral.ai/capabilities/audio_transcription and docs.mistral.ai/getting-started/models, read 2026-11-13. Prices: OpenRouter model pages, 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).
  • 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-voxtral/