JimHue/TASTE-S-Tokenizer

TASTE-S tokenizer

This repository contains the original TASTE-S tokenizer, its shared inference implementation, and the EXP-TOK-021 tokenizer trained on project data. The original model and the new model are kept in separate checkpoint directories; the inference code and dependencies are shared.

Install

git xet install
git clone https://huggingface.co/JimHue/TASTE-S-Tokenizer
cd TASTE-S-Tokenizer
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

A CUDA GPU is recommended for reasonable evaluation time.

Check the original model

The previous repository model is now at checkpoints/taste_s_author_baseline/; it remains the default for inference.py.

python inference.py --wav "samples/audio/*.flac" --out_dir samples/my_output
python samples/check.py samples/my_output

Run either checkpoint

# Original repository model
python inference.py --wav input.wav --out_dir out/original

# Project-data tokenizer (EXP-TOK-021)
python inference.py --model_dir checkpoints/fsq_ourdata_valid_best \
  --wav input.wav --out_dir out/fsq_ourdata

inference.py is shared by both checkpoints. taste_s_tokenizer/ contains the common model implementation; each checkpoint directory contains its own config and weights. campplus.onnx is shared.

Evaluate EXP-TOK-021 on LibriSpeech

See evaluations/README.md. LibriSpeech is downloaded by the evaluator from the OpenSLR LibriSpeech dataset on Hugging Face; no LibriSpeech audio or transcript is stored in this model repository.

The evaluated model is checkpoints/fsq_ourdata_valid_best/. Its upstream task is EXP-TOK-021. The bundled evaluator records CTC-conditioned text WER, reconstructed-audio WER, original-audio HuBERT WER, and UTMOS with per-utterance outputs.

Historical LibriSpeech reference results for this checkpoint:

Metric	EXP-TOK-021 valid_best
CTC-conditioned text WER	5.3810%
Reconstructed-audio WER	4.1645%

These are existing project evaluation results, not a fresh run of the bundled evaluator. They come from 2,620 LibriSpeech test-clean utterances; 2,611 were scored after the protocol's standard exclusion of nine over-length utterances. Use the bundled evaluation instructions to produce a new run for comparison.

Repository organization

• checkpoints/taste_s_author_baseline/ — existing repository checkpoint, relocated without changing its weights.
• checkpoints/fsq_ourdata_valid_best/ — EXP-TOK-021 valid_best.
• inference.py, taste_s_tokenizer/, and campplus.onnx — shared inference and model files.
• evaluations/fsq_ourdata/ — LibriSpeech end-to-end evaluator.
• samples/ — the existing inference examples and references.

Attribution

The model architecture includes Whisper-derived components (MIT) and vendored CosyVoice2/Matcha-TTS code (Apache-2.0/MIT). campplus.onnx is distributed with CosyVoice2. The shipped sample audio is from LibriSpeech; see samples/README.md for sample-specific attribution.