Phonon-2.5

Phonon-2.5 is an open speech recognition model for 25 languages in a 156.2 MB download. On the FLEURS test sets it averages 11.71 % word error across the 25 languages, against 12.18 % for Parakeet Redux, the other low-bit model of the same size, and 13.26 % for Parakeet TDT 0.6B v3, the 2.5 GB full-precision model it derives from. It is ahead of Redux in 14 of the 25 languages. The same file runs on Apple silicon and on Linux and Windows CPUs, with no GPU needed.

For English alone, see Phonon-2.5 English, the same size, English only.

Benchmarks

Word error rate, %, lower is better; the best value in each row is in bold. All three models were scored on the full FLEURS test set of each language with the same harness and the same text normalisation.

Language	Phonon-2.5 (156.2 MB)	Parakeet Redux (178 MB)	Parakeet TDT 0.6B v3 (2,508 MB)
English	6.58	7.32	6.53
Spanish	4.80	4.99	4.69
French	8.12	9.03	6.61
German	6.67	7.57	5.96
Italian	5.21	5.12	4.75
Portuguese	6.53	6.44	6.27
Dutch	9.21	9.45	7.93
Polish	10.20	10.19	8.55
Russian	8.67	9.01	7.21
Ukrainian	7.89	8.05	7.27
Romanian	11.88	11.89	13.13
Czech	10.40	11.41	11.93
Slovak	8.89	8.17	9.72
Bulgarian	13.34	13.02	13.48
Croatian	12.24	11.43	13.21
Swedish	14.53	14.70	15.44
Danish	18.20	17.91	19.05
Finnish	12.02	11.84	13.08
Estonian	11.91	10.29	14.34
Latvian	14.92	14.76	23.66
Lithuanian	17.94	18.05	21.68
Hungarian	15.45	15.81	15.38
Slovenian	18.53	18.88	24.70
Maltese	17.91	16.02	20.86
Greek	20.83	33.25	36.11
Average of 25	11.71	12.18	13.26

On the Open ASR Leaderboard's seven English sets (LibriSpeech clean and other, AMI, Earnings-22, GigaSpeech, SPGISpeech, VoxPopuli, scored with its code on the full test sets) Phonon-2.5 averages 5.06 % word error: 1.60, 3.41, 9.12, 6.72, 8.04, 3.31, 3.23. Phonon-2 averages 5.21 % on the same sets; Phonon-2.5 English averages 4.82 %.

Run it

Python 3.10 or newer (macOS's built-in python3 is 3.9).

pip install fermion-research
fermion transcribe phonon-2.5 recording.wav

On Apple silicon, pip install "fermion-research[mlx]" runs the model on the GPU through MLX. On Linux (x86-64 and Arm) and Windows the same package runs the CPU engine; on Linux, install torch from its CPU wheel index first to skip the GPU build:

pip install --no-deps torch --index-url https://download.pytorch.org/whl/cpu
pip install fermion-research

--json adds a start and end time for every word. fermion serve phonon-2.5 --port 8010 starts an OpenAI-compatible /v1/audio/transcriptions endpoint. The command line and the server are documented at fermionresearch.com/docs/speech.

Notes

Based on parakeet-tdt-0.6b-v3 by NVIDIA; the tokenizer and output conventions (punctuation, casing, numerals) are the original's. Licence CC-BY-4.0, same as the original; NOTICE lists the changes. The command line and this repository's code are Apache 2.0.