| Mach Boom |
|
|
| This repository contains a compressed copy of Phonon 31m (gradium/phonon-31m-en, by Gradium, Apache License 2.0) |
| and four voice embeddings from gradium/phonon-voices (Apache License 2.0), redistributed under the same license. |
| Phonon builds on Pocket TTS and the Mimi codec, developed by Kyutai. |
|
|
| Modifications: the model weights were re-encoded with Mach trellis coding (bitshift trellis, 3 to 4 bits a weight, |
| Hadamard-rotated) and int8 / int4 rounding, and packed into mach-boom.bin; three constant embeddings were folded in at |
| pack time. tokenizer.json is Phonon's tokenizer.5.0.0.json, unchanged. |
|
|
| The runtime code (mach.py, packfmt.py, tmk*.c, tts.py, frontend_en.py) is new code written for this release. |
|
|
