license: cc-by-4.0 base_model: nvidia/parakeet-tdt-0.6b-v3 base_model_relation: quantized language:

en library_name: mlx pipeline_tag: automatic-speech-recognition tags:
mlx
apple-silicon
speech-to-text
asr
stt
on-device metrics:
wer

Phonon-2

Phonon-2 is the most accurate open speech recognition model for English under 900 MB. Across the Open ASR Leaderboard's seven English sets it averages 5.21 % word error, beating models multiple times its size in raw bytes. Set for set it holds the accuracy of the 2.5 GB full-precision parakeet-tdt-0.6b-v3 it derives from, reaching 100.8 % of that model's word accuracy on parliamentary speech and beating it on meetings, from a download 15 times smaller.

It transcribes an hour of audio in about 20 seconds on an M5 MacBook Air (174x realtime), at 143x on eight Zen 5 cores (16 vCPU) and at 6,680x on one H100 in batches of 128.

Benchmarks

Model	Download	LS clean	LS other	AMI	Earnings-22	GigaSpeech	SPGISpeech	VoxPopuli	Average
Phonon-2	164 MB	1.72	3.92	9.37	6.96	8.35	3.70	2.46	5.21
Parakeet TDT 0.6B v3, base model†	2,508 MB	1.52	3.13	9.42	5.85	7.99	3.63	3.19	4.96
Parakeet Redux	178 MB	1.94	4.35	9.16	7.90	8.62	4.01	3.87	5.69
Phonon-1	415 MB	2.11	5.03	10.31	12.34	8.73	3.67	3.73	6.56
Canary 180M Flash†	737 MB	1.52	3.42	12.09	8.33	8.87	2.04	3.57	5.69
Voxtral Mini 4B Realtime†	≈8,000 MB*	1.62	4.94	13.34	9.31	8.80	2.23	2.60	6.12
Whisper large-v3-turbo†	1,618 MB	2.13	3.71	13.88	8.09	8.47	2.79	7.02	6.58
Nemotron 3.5 ASR Streaming 0.6B†	2,368 MB	2.83	6.79	13.43	15.30	9.86	3.27	4.24	7.96

† Open ASR Leaderboard's published row; the other rows use its code on the full test sets. * Size from the parameter count at 16 bits.

Run it

Phonon-2 is the model inside Detta, the dictation app for the Mac.

On Apple silicon, from the command line (Python 3.10 or newer; macOS's built-in python3 is 3.9):

python3.12 -m venv .venv && source .venv/bin/activate
pip install fermion-research
pip install mlx mlx-audio mlx-lm soundfile scipy zstandard
phonon transcribe recording.wav
fermion transcribe phonon-2 recording.wav

--json adds a start and end time for every word. The command line and the server are documented at fermionresearch.com/docs/speech.

Core ML (Neural Engine): huggingface.co/FermionResearch/Phonon-2-CoreML

Android (arm64-v8a, API 26+): the CPU engine libraries, with the C header and a note on the call sequence: https://github.com/fermionresearch/phonon/releases/tag/android-arm64-v8a-0.2.11

The same package runs on Linux (x86-64 and Arm) and Windows CPUs. With Docker, on a CPU or a GPU:

docker run --rm -v "$PWD":/audio -v phonon-cache:/home/phonon/.cache ghcr.io/fermionresearch/phonon-cpu:2.0.7 transcribe phonon-2 /audio/recording.wav
docker run --rm --gpus all -v "$PWD":/audio -v phonon-cache:/home/phonon/.cache ghcr.io/fermionresearch/phonon-cuda:1.0.6 transcribe phonon-2 /audio/recording.wav

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.