Fermion ResearchFermion Research

Phonon/facts checked 21 November 2026

The Open ASR Leaderboard and the Phonon models

The leaderboard is the public reference for English speech recognition accuracy. This page explains its protocol and lists the Phonon numbers scored with its code.

Open ASR Leaderboard on Hugging Face

What it measures

The Open ASR Leaderboard, run by Hugging Face, scores speech models on public English test sets (LibriSpeech clean and other, AMI, Earnings-22, GigaSpeech, SPGISpeech, VoxPopuli, and newer additions) and reports word error rate per set, an average, and RTFx, the inverse realtime factor. The maintainers run every model with one evaluation harness and one text normaliser, so rows are comparable.

Submissions are pull requests to the open_asr_leaderboard repository with an evaluation script; the maintainers run the official job.

Phonon rows

  • Phonon-2.5 English (158 MB) is listed with a board average of 7.05 % and RTFx 2,110. Our seven-set average with the leaderboard's code is 4.82 %.
  • Phonon-2.5 (156 MB, 25 languages) is listed with a board average of 7.16 % and RTFx 1,898. Our seven-set average is 5.06 %.
  • Phonon-2 (164 MB, English) had a row until 2026-11-20, when the board removed it. Our seven-set average is 5.21 %.

The board's printed average can differ from a seven-set average because the board includes additional sets; the numbers here state which sets they cover.

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

Normalisation

The leaderboard applies the Whisper English text normaliser before scoring: case, punctuation, number and spelling variants are folded, so a transcript is not penalised for writing 'twenty' where the reference has '20'. Multilingual rows use a different normaliser that keeps diacritics, which is why FLEURS numbers on the Phonon-2.5 page, scored with our own harness, are not the same thing as leaderboard cells.

Questions

Is Phonon on the Open ASR Leaderboard?
Yes. Phonon-2.5 English and Phonon-2.5 are listed since 2026-11-13, at 7.05 % and 7.16 % on the board's default average.
What is Phonon-2.5 English's word error rate?
4.82 % averaged over the seven classic Open ASR Leaderboard English sets, scored with the leaderboard's code on the full test sets.
What does RTFx mean?
Inverse realtime factor: audio seconds transcribed per second of compute. 1,000 RTFx means an hour of audio in 3.6 seconds on the maintainers' hardware.
Why does the board number differ from the model card?
The board's default average leaves out LibriSpeech, adds URGENT24, Monsoon en-IN and two private sets, and runs on its own hardware. The card's 4.82 % and 5.06 % cover the seven classic sets.

Related

Sources and method

  • Leaderboard protocol: github.com/huggingface/open_asr_leaderboard README, read 2026-11-13.
  • Phonon numbers: Phonon-2.5 release post Table 2 (2026-10-22). Board rows: Open ASR Leaderboard version 1.16.2, read 2026-11-21 (Phonon-2.5 added in version 1.16.0 on 2026-11-13; Phonon-2 removed in 1.16.2 on 2026-11-20).

Canonical: https://www.fermionresearch.com/models/phonon-2-5/open-asr-leaderboard/