Fermion ResearchFermion Research

Speech recognition/facts checked 21 November 2026

English speech recognition benchmark results

Per-set word error rates for Phonon-2.5 English, Phonon-2.5 and Phonon-2 on the Open ASR Leaderboard's English test sets, with the leaderboard's own rows for context.

Phonon-2.5 English158 MB, open weights on Hugging Face.

The short answer

On the seven classic English sets of the Open ASR Leaderboard, scored with the leaderboard's code on the full test sets, Phonon-2.5 English averages 4.82 % word error, the 25-language Phonon-2.5 5.06 % and Phonon-2 5.21 %.

On the nine public sets of the leaderboard's current protocol, Phonon-2.5 English scores 5.60 % and Phonon-2.5 5.70 % (Fermion Research run of 2026-10-28). Phonon-2's leaderboard row on the same nine sets was 6.25 %.

Phonon-2.5 English and Phonon-2.5 are listed on the Open ASR Leaderboard since 2026-11-13 (board version 1.16.0), with board averages of 7.05 % and 7.16 % over its default columns (seven public English sets, without LibriSpeech, plus two private sets the maintainers run) and RTFx of 2,110 and 1,898 on the maintainers' GPU. The board removed Phonon-2's row on 2026-11-20.

Seven English sets

Word error rate, %, lower is better. † marks the Open ASR Leaderboard's published row; the other rows were scored with the leaderboard's code on the full test sets, with Earnings-22 as the leaderboard's chunked set. Table as published in the Phonon-2.5 release post on 2026-10-22.

Open ASR Leaderboard, seven English sets
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 v3 †2,508 MB1.523.139.425.857.993.633.194.96
Parakeet Redux178 MB1.944.359.167.908.624.013.875.69

Nine English sets (current protocol)

The leaderboard's current English protocol adds URGENT24 and Monsoon en-IN to the seven sets. Fermion Research ran it with the leaderboard's own code on the full sets on 2026-10-28. Both models are now on the board, which since November 2026 leaves LibriSpeech out of its default average and adds two private sets, so the board's averages (7.05 % and 7.16 %) cover different sets from the table below. Phonon-2's average on the same nine sets was its leaderboard row, 6.25 %.

The multilingual Phonon-2.5 is slightly lower than Phonon-2.5 English on URGENT24, Monsoon en-IN, both LibriSpeech sets and SPGISpeech; Phonon-2.5 English is lower on AMI, Earnings-22, GigaSpeech and VoxPopuli, and on the average. URGENT24, speech with noise and distortions, is the hardest set for both.

Fermion Research run, 2026-10-28, Open ASR Leaderboard code, full test sets
ModelAMIEarnings-22GigaSpeechLS cleanLS otherSPGISpeechURGENT24Monsoon en-INVoxPopuliAverage
Phonon-2.5 English (158 MB)8.696.187.991.643.563.3212.324.442.305.60
Phonon-2.5 (156 MB, 25 languages)8.996.688.011.613.493.3011.884.293.015.70

Context from the leaderboard

Other models' seven-set averages from the Open ASR Leaderboard snapshot of 2026-09-25, for comparison with the seven-set numbers above only. Several larger open models and closed APIs have lower averages than Phonon-2.5 English; among models under 200 MB on that snapshot, Phonon-2 had the lowest average.

Open ASR Leaderboard, seven English sets, snapshot 2026-09-25
ModelSizeSeven-set average WER %
zoom/scribe_v2_pro (closed API)n/a3.68
elevenlabs/scribe_v2 (closed API)n/a4.06
HojoAI/Hojo-ASR-V110,360 MB4.29
Qwen/Qwen3-ASR-1.7B-hf4,080 MB4.43
nvidia/parakeet-tdt-0.6b-v22,472 MB4.82
nvidia/parakeet-tdt-0.6b-v32,508 MB4.96
Phonon-2164 MB5.21
openai/whisper-large-v33,090 MB6.04
openai/whisper-large-v3-turbo1,618 MB6.58
abr-ai/niagara-84m-batch.en168 MB7.27

Questions

What is Phonon-2.5 English's word error rate?
4.82 % averaged over the seven classic Open ASR Leaderboard English sets, and 5.60 % over the nine sets of the current protocol, both scored with the leaderboard's code on the full test sets.
Is Phonon-2.5 on the Open ASR Leaderboard?
Yes, since 2026-11-13. The board averages 7.05 % for Phonon-2.5 English and 7.16 % for Phonon-2.5 over its default columns, which leave out LibriSpeech and include two private sets.
Is the 25-language Phonon-2.5 good at English?
It averages 5.06 % on the seven sets, lower than Phonon-2's 5.21 %. Phonon-2.5 English is lower still at 4.82 %.
Which English set is hardest for Phonon?
URGENT24, speech with noise and distortions: 11.88 % for Phonon-2.5 and 12.32 % for Phonon-2.5 English. Among the classic seven, AMI meeting recordings is the hardest.
Do any models score lower than Phonon-2.5 English?
Yes. On the 2026-09-25 leaderboard snapshot, several larger open models and closed APIs have lower seven-set averages, the lowest being 3.68 % for a closed API.
Why do the seven-set and nine-set averages differ?
They average different sets. The nine-set protocol adds URGENT24 and Monsoon en-IN, and URGENT24's noisy audio raises the average.

Related

Sources and method

  • Seven-set English numbers: Phonon-2.5 release post Table 2, 2026-10-22; Open ASR Leaderboard scoring code on the full test sets (Earnings-22 = the leaderboard's chunked set); † = the Open ASR Leaderboard's published row.
  • Nine-set English numbers: Fermion Research measurement, 2026-10-28, Open ASR Leaderboard code, full test sets; Phonon-2 = its leaderboard row (removed from the board 2026-11-20). Official board rows: Open ASR Leaderboard version 1.16.2, read 2026-11-21.
  • Other models, seven English sets: Open ASR Leaderboard, seven English sets, snapshot 2026-09-25.

Canonical: https://www.fermionresearch.com/speech/english-benchmark-results/