The short answer
Choose Whisper when you need one of the languages Phonon-2.5 does not cover: Whisper's model card lists 99 languages, Phonon-2.5 covers 25, all European. Choose Phonon-2.5 when the language is one of the 25 and you want a small download that runs on a laptop CPU or Apple silicon with no GPU.
On the FLEURS test sets, Phonon-2.5 averages 11.71 % word error across its 25 languages. That average is lower than Whisper large-v3, a model about 20 times its size, and lower than Whisper large-v3-turbo, about 10 times its size. Whisper base, which is about the same size as Phonon-2.5, averages 53.71 % on the same 25 sets, 4.6 times as many errors.
For English alone, Phonon-2.5 English (158 MB) averages 4.82 % word error on the seven English test sets of the Open ASR Leaderboard, and the 25-language Phonon-2.5 averages 5.06 % on the same sets.
Side by side
Facts about Whisper are from OpenAI's whisper repository and model card, read 2026-11-13. Phonon numbers are from the Phonon-2.5 release post.
| Phonon-2.5 | Whisper large-v3 | Whisper base | |
|---|---|---|---|
| Download | 156 MB | about 3 GB (1.55 B parameters) | about 150 MB (74 M parameters) |
| Languages | 25 | 99 | 99 |
| FLEURS word error, average of Phonon-2.5's 25 languages | 11.71 % | higher than Phonon-2.5 (release post) | 53.71 % |
| Seven-set English word error (Open ASR Leaderboard sets) | 5.06 % (Phonon-2.5 English: 4.82 %) | not scored by us | not scored by us |
| Runs on | CPU, Apple silicon (MLX, Neural Engine via Core ML for Phonon-2), NVIDIA GPUs, Android | GPU, or CPU through whisper.cpp | GPU or CPU |
| Install | pip install fermion-research | pip install openai-whisper, or whisper.cpp | same as large-v3 |
Where Whisper is the better choice
- A language outside Phonon-2.5's 25: Whisper covers Arabic, Chinese, Hindi, Japanese, Korean, Turkish and dozens more.
- Translation to English: Whisper can translate speech to English text; Phonon-2.5 transcribes only.
- An existing whisper.cpp or faster-whisper pipeline you do not want to change.
Where Phonon-2.5 is the better choice
- Size: 156 MB for 25 languages. A phone, a laptop without a GPU, or a container with a small image.
- Accuracy per megabyte: a lower FLEURS average than Whisper large-v3 from a twentieth of the download.
- Apple silicon: Phonon runs on the MLX engine, and Phonon-2 runs on the Neural Engine through Core ML at 606 times realtime on an M5 MacBook Air (Phonon-2's published figure).
- A server with an OpenAI-compatible endpoint from one command: fermion serve phonon-2.5.
pip install fermion-research
fermion transcribe phonon-2.5 meeting.wav
fermion serve phonon-2.5 # POST /v1/audio/transcriptionsSpeed
Phonon-2.5 has the architecture and size of Phonon-2, and its published speed figures are Phonon-2's: 174 times realtime on the GPU of a base M5 MacBook Air (an hour of audio in about 20 seconds), 143 times realtime on eight AMD Zen 5 cores, and 6,680 times realtime batched on an H100. Measured Phonon-2.5 speed figures will be published when they exist.
Questions
- Is Phonon-2.5 more accurate than Whisper?
- On the FLEURS test sets for its 25 languages, Phonon-2.5 averages 11.71 % word error, which is lower than Whisper large-v3 and Whisper large-v3-turbo on the same sets, and far lower than Whisper base at 53.71 %. Whisper covers 74 more languages that Phonon-2.5 cannot transcribe at all.
- How big is Phonon-2.5 compared with Whisper?
- Phonon-2.5 is a 156 MB download (178 MB on disk). Whisper large-v3 is about 3 GB and Whisper base about 150 MB.
- Which languages does Phonon-2.5 support?
- Twenty-five: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Romanian, Czech, Slovak, Bulgarian, Croatian, Swedish, Danish, Finnish, Estonian, Latvian, Lithuanian, Hungarian, Slovenian, Maltese and Greek.
- Can Phonon-2.5 run without a GPU?
- Yes. It runs on a laptop CPU, on Apple silicon through MLX, on NVIDIA GPUs, and on Android. Phonon-2 also runs on the Apple Neural Engine through Core ML.
- Is Phonon-2.5 a drop-in replacement for the Whisper API?
- fermion serve phonon-2.5 exposes POST /v1/audio/transcriptions, the same route shape as OpenAI's transcription endpoint, so a client that posts a file to that route works against it. Whisper-specific options such as translation are not offered.
- Where are the Phonon-2.5 weights?
- On Hugging Face under FermionResearch/Phonon-2.5, and the fermion command downloads them on first use.
Related
Sources and method
- Phonon-2.5 FLEURS and English numbers: the Phonon-2.5 release post (/research/phonon-2-5/), full test sets, one harness for every model in the FLEURS table.
- Whisper sizes, parameter counts and language count: github.com/openai/whisper README and model card, read 2026-11-13.
- Speed: Phonon-2's published figures (/models/phonon-2/); no Phonon-2.5 speed figures are published.
Canonical: https://www.fermionresearch.com/models/phonon-2-5/vs-whisper/