use chatterbox MTLTokenizer for multilingual.

[litmudoc]

Context

The Chatterbox TTS model supports multilingual text-to-speech, but the tokenizer selection code was not properly configuring MTLTokenizer for multilingual models. When a model with "multilingual": true in its config..json was loaded, it would incorrectly use the English tokenizer (EnTokenizer) instead of the multilingual tokenizer.

Description

Modified _init_tokenizers() and from_pretrained(), post_load_hook() methods in chatterbox.py to:

1. Check for the existence of config.json in the model directory
2. Read the "multilingual" configuration flag from JSON config
3. Select and instantiate MTLTokenizer when multilingual is enabled, or EnTokenizer for English-only models
4. Added appropriate console logging messages indicating which tokenizer was loaded

Changes in the codebase

• Tokenizer Selection Logic: Added conditional logic to check config.["multilingual"] and instantiate the appropriate tokenizer class
• Import Updates: Added MTLTokenizer to imports alongside existing EnTokenizer
• User Feedback: Added print statements showing which tokenizer was loaded ("Loaded multilingual tokenizer (MTLTokenizer)" or "Loaded English tokenizer (EnTokenizer)")
• Three Locations Modified: _init_tokenizers(), load() method for S3 checkpoints, and load() method for model loading

Additional information

This change enables Chatterbox multilingual TTS models to properly tokenize input text, which is required for text-to-speech generation in multiple languages. The implementation follows the existing pattern of checking model configuration files.

Checklist

• Code tested with multilingual Chatterbox models
• Documentation updated for multilingual model usage
• No breaking changes to existing English-only models

[litmudoc]

I tried to use MTLTokenizer for Chatterbox Multilingual support. Thank you.

[Blaizzy]

Hey @litmudoc

Awesome, do you have any audio samples .?

[litmudoc]

mlx_audio.tts.generate \
    --model litmudoc/Chatterbox-Multilingual-MLX-v2-fp16 \
    --exaggeration 0.6 \
    --cfg_scale 0.35 \
    --temperature 0.8 \
    --lang_code ko \
    --text ", 한국말이 너무 자연스러워요\! 감격하고, 또 감격 했습니다." \
    --ref_audio ko.wav \
    --ref_text "우리는 정말로 허름한 호텔에 묵었지만, 그래도 행복했다." \
    --verbose --play

This is a sample created with the Chatterbox multilingual model on local Mac.
audio_000.wav

• I made a sample using the Korean wav obtained from the Chatterbox Multilingual Demo. I'm very pleased!
  Chatterbox Multilingual Demo
• Thanks to your amazing mlx_audio, the Chatterbox Multilingual model is running smoothly on my local Mac. Thank you!
  Chatterbox-Multilingual-MLX-v2-fp16

[Blaizzy]

LGTM, thanks!