training run: tuning ichigo-quantizer

[hahuyhoang411]

Goal

Tuning to find the best params for Ichigo Quantizer

Hypothesis

• Current default codebook (512) is too small for multilingual
• Training on mixed dataset (Eng-Vie) maintains the performance of original Whisper

Task:

• [x] Adding code to init new codebook embedding with avg.
• [ ] Adding code to preprocess multi training datasets

[hahuyhoang411]

The training result here: https://github.com/janhq/WhisperSpeech/issues/2

[hahuyhoang411]

GJ @tuanlda78202 on adding code to utilize weight from whispervq 7lang quantizer checkpoint PR: https://github.com/janhq/WhisperSpeech/pull/7

This could cut off plenty of time instead of train from scratch