Fermion ResearchFermion Research

Phonon/facts checked 13 November 2026

Phonon-2.5 vs NVIDIA Canary

NVIDIA's Canary family spans an English-only 2.5B model and a 25-language 1B model. Here is how each compares with Phonon-2.5 on English, languages, size and where they run.

Phonon-2.5 downloads and benchmarks

The short answer

Canary-1B-v2 covers exactly the same 25 European languages as Phonon-2.5 and adds speech translation to and from English; Phonon-2.5 has the lower English word error (5.06 % against 5.81 % on the Open ASR Leaderboard's seven sets) from 156 MB against 3,916 MB. For English alone, Canary-Qwen 2.5B is lower than both Phonon models (4.55 % against 4.82 % for Phonon-2.5 English).

Choose Canary for translation, for the lowest English word error in an NVIDIA GPU stack, or if you already run NeMo. Choose Phonon-2.5 for the same 25 languages with no GPU and a twenty-fifth of the download.

Side by side

Canary facts from NVIDIA's model cards on Hugging Face, read 2026-11-13. Sizes from the leaderboard snapshot.

Phonon-2.5Canary-1B-v2Canary-Qwen 2.5B
Model typeOpen weights on Hugging FaceOpen weightsOpen weights
Size156 MB download (Phonon-2.5 English: 158 MB)978 M parameters, 3,916 MB2.5 B parameters, 5,000 MB
Languages25 European languages (Phonon-2.5 English: English only)The same 25 European languagesEnglish
English word error, seven Open ASR sets5.06 % (Phonon-2.5 English: 4.82 %)5.81 %4.55 %
PriceNo per-hour fee when self-hostedFree to run; no hosted price checkedFree to run; no hosted price checked
Where it runsYour hardware: CPUs (Linux x86-64 and arm64, Windows x86-64), Apple silicon, NVIDIA GPUs, AndroidNVIDIA GPUs on Linux with NeMoNVIDIA GPUs with NeMo (Linux, Windows)
StreamingYes, live WebSocket from fermion serveNot mentioned on the cardNot mentioned on the card
Speech translationNoEnglish to 24 languages and 24 to EnglishNo (LLM mode answers questions about the transcript)
Word timestampsYes (--json output)Word and segment levelNot mentioned on the card

English, set by set

Canary-Qwen 2.5B is lower than Phonon-2.5 on six sets; Phonon-2.5 English is lower on Earnings-22 and VoxPopuli. Canary-1B-v2 is lower than Phonon-2.5 on LibriSpeech other and SPGISpeech and higher on the rest. Two smaller English Canary models are also in the snapshot: Canary-1B-Flash at 5.03 % (3,244 MB) and Canary-180M-Flash at 5.69 % (737 MB).

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 EnglishCanary-1B-v2Canary-Qwen 2.5BCanary-1B-FlashCanary-180M-Flash
LibriSpeech clean1.601.641.761.231.201.52
LibriSpeech other3.413.563.072.632.523.42
AMI9.128.7613.037.9110.5812.09
Earnings-226.726.227.286.416.798.33
GigaSpeech8.047.999.127.808.198.87
SPGISpeech3.313.312.031.701.752.04
VoxPopuli3.232.284.394.144.173.57
Average of the seven5.064.825.814.555.035.69

The 25 languages

Canary-1B-v2's card reports 8.40 % word error on FLEURS across its 25 languages, measured without punctuation and capitalisation. Phonon-2.5's 11.71 % FLEURS average comes from our own scoring code, and we have not scored Canary-1B-v2 with it, so the two figures should not be set against each other. If one language matters most, test both on your own audio in that language.

Where each is the better choice

  • Canary: speech translation in 48 directions, English accuracy with Canary-Qwen 2.5B, an LLM mode that summarises or answers questions about a transcript, and NeMo pipelines on NVIDIA GPUs.
  • Phonon-2.5: the same 25 languages from 156 MB, CPUs, Apple silicon and Android with no GPU, lower English word error than Canary-1B-v2, and a ready server with file and streaming routes.
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 NVIDIA Canary?
Than Canary-1B-v2 in English, yes: 5.06 % against 5.81 % on the Open ASR Leaderboard's seven sets. Canary-Qwen 2.5B (4.55 %) and Canary-1B-Flash (5.03 %) are lower than Phonon-2.5; Phonon-2.5 English (4.82 %) is lower than Canary-1B-Flash but not Canary-Qwen 2.5B.
Do Canary-1B-v2 and Phonon-2.5 cover the same languages?
Yes. NVIDIA's card lists the same 25 European languages Phonon-2.5 covers, from Bulgarian to Ukrainian, including Maltese and Greek.
Can Phonon-2.5 translate like Canary?
No. Canary-1B-v2 translates speech from English into 24 languages and from 24 languages into English; Phonon-2.5 only transcribes.
Does Canary run without an NVIDIA GPU?
NVIDIA's cards list NVIDIA GPU architectures and the NeMo toolkit. Phonon-2.5 runs on CPUs, Apple silicon, NVIDIA GPUs and Android.
How much smaller is Phonon-2.5 than Canary?
Phonon-2.5 is 156 MB. In the leaderboard snapshot Canary-1B-v2 is 3,916 MB, Canary-Qwen 2.5B 5,000 MB and Canary-180M-Flash 737 MB.

Related

Sources and method

  • Canary-1B-v2 languages, translation, timestamps, runtime and FLEURS figure: huggingface.co/nvidia/canary-1b-v2, read 2026-11-13.
  • Canary-Qwen 2.5B parameters, language, modes and runtime: huggingface.co/nvidia/canary-qwen-2.5b, read 2026-11-13.
  • Rival English word error: Open ASR Leaderboard, seven English sets, snapshot 2026-09-25 (huggingface.co/spaces/hf-audio/open_asr_leaderboard). Sizes from the same snapshot.
  • 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.
  • Phonon-2.5 FLEURS average: Phonon-2.5 release post, full FLEURS test sets, our scoring code.

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