Docs / Speech
Speech models
Every released Phonon repository, its sizes, and its measured word error rates.
Phonon-1
The default speech recognition model and the one behind fermion transcribe, listen, and speech serving. It transcribes English from 16 kHz mono audio and was trained at 2.4 bits per weight from the start. The aliases phonon, stt, and asr all resolve to it. Benchmark figures are word error rate; lower is better.
| Download | On disk | LibriSpeech test-clean | LibriSpeech test-other | Eight-benchmark macro |
|---|---|---|---|---|
| 415 MB | 455 MB | 2.64 | 5.70 | 7.67 |
Phonon-1 Micro
The smallest speech model. It trades accuracy for a smaller download and is selected with --model micro.
| Download | On disk | LibriSpeech test-clean | LibriSpeech test-other | Eight-benchmark macro |
|---|---|---|---|---|
| 285 MB | 331 MB | 3.00 | 6.51 | 8.52 |
Phonon-1-Big
| Download | On disk | LibriSpeech test-clean | LibriSpeech test-other | Eight-benchmark macro |
|---|---|---|---|---|
| 581 MB | 822 MB | 2.67 | 5.72 | 7.60 |