google / lyra

A Very Low-Bitrate Codec for Speech Compression
Apache License 2.0
3.84k stars 355 forks source link

Update models #104

Closed mchinen closed 2 years ago

mchinen commented 2 years ago

The main change is the new tflite models, which use 8-bit quantization via quantization-aware training to obtain the same quality with a significant reduction in complexity (20% faster) and space (43% smaller). The numeric version is 1.3.0. This change breaks the bitstream, so this model is not compatible with with output of previous models. The other changes are primarily refactoring and documentation.

Change-Id: I273606b5a77ef52ba80191d7b8a10c4905046f63