Fermion ResearchFermion Research

Phonon/facts checked 13 November 2026

Open speech-to-text models under 200 MB, compared

If the model has to fit on a phone, in a small container or in a laptop app, the question is accuracy per megabyte. These are the published reads for the open models in the 150 to 200 MB class.

Phonon-2.5 downloads and benchmarks

English: seven Open ASR Leaderboard sets

Word error rate in %, lower is better. Scored with the leaderboard's code on the full test sets (Earnings-22 is the leaderboard's chunked set). The Parakeet TDT 0.6B v3 row is a 2.5 GB model, included for scale.

ModelSizeLS cleanLS otherAMIEarnings-22GigaSpeechSPGISpeechVoxPopuliAverage
Phonon-2.5 English158 MB1.643.568.766.227.993.312.284.82
Phonon-2.5 (25 languages)156 MB1.603.419.126.728.043.313.235.06
Phonon-2164 MB1.723.929.376.968.353.702.465.21
Underdog Bark 3.0163 MB1.894.087.827.057.953.722.054.94
Parakeet TDT 0.6B v32,508 MB1.523.139.425.857.993.633.194.96
Parakeet Redux178 MB1.944.359.167.908.624.013.875.69

25 languages: FLEURS

Two models in this class cover 25 languages. Phonon-2.5 (156 MB) averages 11.71 % and Parakeet Redux (178 MB) 12.18 % on the full FLEURS test sets; the 2.5 GB Parakeet TDT 0.6B v3 averages 13.26 %. Whisper base, about 150 MB and 99 languages, averages 53.71 % on the same 25 sets. The per-language table is on the Phonon-2.5 vs Parakeet page.

How to read it

  • English only: Phonon-2.5 English, 4.82 %, is the lowest average in the class on these seven sets; Underdog Bark 3.0 is ahead on AMI (meetings), GigaSpeech and VoxPopuli, so meeting-heavy audio deserves a test on your own recordings.
  • Several European languages: Phonon-2.5, the only model under 160 MB with 25 languages and a FLEURS average below the 2.5 GB base.
  • Models not in the table: Moonshine (English, tiny and base), Whisper tiny/base/small and Vosk models were not scored by us with this protocol, so no number is given for them here.

Run any Phonon model

One package covers CPU, Apple silicon and NVIDIA GPUs; the same command serves an OpenAI-compatible endpoint.

pip install fermion-research
fermion transcribe phonon-2.5-en call.wav
fermion transcribe phonon-2.5 meeting.wav --json
fermion serve phonon-2.5-en

Questions

What is the most accurate open English speech model under 200 MB?
On the seven Open ASR Leaderboard sets scored with the leaderboard's code, Phonon-2.5 English (158 MB) averages 4.82 % word error, ahead of Underdog Bark 3.0 (163 MB, 4.94 %), Phonon-2 (164 MB, 5.21 %) and Parakeet Redux (178 MB, 5.69 %).
What is the most accurate open multilingual speech model under 200 MB?
Among models we scored, Phonon-2.5 (156 MB, 25 languages) at 11.71 % average FLEURS word error, against 12.18 % for Parakeet Redux (178 MB) and 53.71 % for Whisper base.
Why compare on FLEURS and the Open ASR Leaderboard sets?
They are public test sets with published protocols, so anyone can re-run the numbers. Every model in the tables was scored on the full sets with the same code.
Does a smaller model mean faster transcription?
Usually, because less memory moves per step. Phonon-2's published speed is 174 times realtime on an M5 MacBook Air GPU and 143 times on eight AMD Zen 5 cores; Phonon-2.5 shares its size and architecture.

Related

Sources and method

  • English table: Phonon-2.5 release post Table 2 (2026-10-22), leaderboard scorer, full sets; Parakeet TDT 0.6B v3 is its published row.
  • FLEURS averages and Whisper base: Phonon-2.5 release post, 2026-10-22.
  • Speed: Phonon-2 model page, published figures.

Canonical: https://www.fermionresearch.com/models/phonon-2-5/small-models/