BanglaBox — Bangla Chatterbox TTS with zero-shot voice cloning
Anonymous artifact for a double-blind submission. Everything needed to reproduce the paper's Bangla
text-to-speech system is in this repository: the fine-tuned adapter, the extended tokenizer, the text
normalisation and the complete training and
inference code. The evaluation materials (test sentences, generated audio for every system, per-clip
scores and the scoring scripts) are in the companion dataset
Banglabox/banglaeval; the speech corpus is Banglabox/bangla-corpus.
Repository layout
checkpoint/
NEW_VOCAB_SIZE.txt
adapter/adapter_config.json # LoRA configuration
adapter/adapter_model.safetensors
inference/
infer.py # BanglaTTS class + CLI (locked inference used for every reported number)
bn_norm.py # Stage-2 normalisation (for_tts) and CER reference normalisation (for_cer)
fix_bangla_graphemes.py # grapheme-repair utility
NEW_VOCAB_SIZE.txt, requirements.txt
src/ # Chatterbox model code (T3 / S3Gen / voice encoder) with the Bangla changes
training/
add_bangla_tokens.py # Stage-1 tokenizer extension (merge-consistent filter)
train.py, train_modal.py # fine-tuning entry points (HF Trainer)
setup.py, requirements.txt
src/config.py # all hyper-parameters (Table "Fine-tuning and inference configuration")
src/model.py # vocabulary resize + mean-init, prompt-masked dual loss
src/dataset.py, src/preprocess_*.py
pretrained_models/ # base Chatterbox: t3_cfg.safetensors, s3gen.safetensors, ve.safetensors, conds.pt, tokenizer.json
voices/ # example reference clips (reference_merged.wav is the reference used for all reported SECS)