Fermion ResearchFermion Research

Phonon/facts checked 13 November 2026

Phonon-2.5 vs Whisper large-v3-turbo

Two open speech models you can run yourself. Whisper large-v3-turbo is OpenAI's faster Whisper with 99 languages; Phonon-2.5 covers 25 European languages from a tenth of the download.

Phonon-2.5 downloads and benchmarks

The short answer

Pick Phonon-2.5 when your audio is in one of its 25 European languages: on the seven English test sets of the Open ASR Leaderboard it averages 5.06 % word error against 6.58 % for Whisper large-v3-turbo, from a 156 MB download against 1,618 MB. Pick Whisper large-v3-turbo when you need a language outside those 25 (its model card lists 99) or speech translation into English.

Across Phonon-2.5's 25 languages on the FLEURS test sets, the Phonon-2.5 release post reports a lower average word error for Phonon-2.5 than for Whisper large-v3-turbo. Phonon-2.5 English, the English-only model, averages 4.82 % on the seven English sets.

Side by side

Whisper facts from OpenAI's model card on Hugging Face, read 2026-11-13. Hosted prices are list prices from third-party hosts; OpenAI publishes the weights.

Phonon-2.5Whisper large-v3-turbo
Model typeOpen weights on Hugging FaceOpen weights on Hugging Face
Size156 MB download (Phonon-2.5 English: 158 MB)809 M parameters, 1,618 MB
Languages25 European languages (Phonon-2.5 English: English only)99
English word error, seven Open ASR sets5.06 % (Phonon-2.5 English: 4.82 %)6.58 %
Hosted price per audio hourNo per-hour fee when self-hostedDeepInfra $0.012; Groq $0.04 (third-party hosts)
Where it runsYour hardware: CPUs (Linux x86-64 and arm64, Windows x86-64), Apple silicon, NVIDIA GPUs, AndroidGPU or CPU through Transformers, whisper.cpp and other runtimes
StreamingYes, live WebSocket from fermion serveChunked long-form decoding in Transformers
Speech translation to EnglishNoYes
Word timestampsYes (--json output)Yes (return_timestamps="word")

English, set by set

Phonon-2.5 has the lower word error on six of the seven sets. Whisper large-v3-turbo is ahead on SPGISpeech (financial dictation), 2.79 % against 3.31 %. The widest gaps are AMI meetings (13.88 % against 9.12 %) and VoxPopuli parliamentary speech (7.02 % against 3.23 %).

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 EnglishWhisper large-v3-turbo
LibriSpeech clean1.601.642.13
LibriSpeech other3.413.563.71
AMI9.128.7613.88
Earnings-226.726.228.09
GigaSpeech8.047.998.47
SPGISpeech3.313.312.79
VoxPopuli3.232.287.02
Average of the seven5.064.826.58

Where Whisper large-v3-turbo is the better choice

  • Languages outside Europe: Chinese, Japanese, Arabic, Hindi, Korean, Turkish and the rest of Whisper's 99.
  • Translation: setting the task to "translate" produces English text from speech in another language. Phonon transcribes only.
  • Existing Whisper tooling: whisper.cpp, faster-whisper and many hosted endpoints already serve it; DeepInfra lists it at $0.00020 per minute ($0.012 per audio hour).
  • Financial dictation in English: its SPGISpeech score is lower than Phonon-2.5's.

Where Phonon-2.5 is the better choice

  • English accuracy on meetings, earnings calls, podcasts and parliamentary speech: lower word error on six of seven sets.
  • Size: 156 MB fits a phone, a laptop app or a small container; turbo is 1,618 MB.
  • Live transcription: fermion serve exposes a WebSocket stream next to the file route.
  • Verbatim output: Phonon keeps repetitions and short responses such as "Mm-hmm" as spoken.
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 Whisper large-v3-turbo?
In English, yes on the Open ASR Leaderboard's seven sets: 5.06 % average word error against 6.58 %, lower on six of seven sets. Whisper large-v3-turbo is lower on SPGISpeech, and it covers 74 languages Phonon-2.5 does not.
How much smaller is Phonon-2.5 than Whisper large-v3-turbo?
Phonon-2.5 is a 156 MB download. Whisper large-v3-turbo has 809 million parameters and is 1,618 MB, about ten times larger.
Can Phonon-2.5 translate speech like Whisper?
No. Phonon-2.5 transcribes speech in the language it was spoken. Whisper large-v3-turbo can translate speech into English text.
Do both models give word timestamps?
Yes. Phonon prints word timings with the --json flag; Whisper large-v3-turbo returns them in Transformers with return_timestamps set to "word".
Which is better for European languages?
If the language is one of Phonon-2.5's 25, the Phonon-2.5 release post reports a lower FLEURS average for Phonon-2.5 than for Whisper large-v3-turbo across those languages. For languages outside the 25, use Whisper.

Related

Sources and method

  • Whisper large-v3-turbo parameters, languages, translation and timestamps: huggingface.co/openai/whisper-large-v3-turbo, read 2026-11-13.
  • DeepInfra price: deepinfra.com/openai/whisper-large-v3-turbo, read 2026-11-13. Groq price: 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.
  • FLEURS comparison: Phonon-2.5 release post (/research/phonon-2-5/), full FLEURS test sets, one scoring code for all models.

Canonical: https://www.fermionresearch.com/models/phonon-2-5/vs-whisper-large-v3-turbo/