accentbox-zh

AccentBox Chinese

Chinese AccentBox checkpoint for multi-accent TTS with independent speaker and accent references.

Model

• Acoustic model: AccentBox Chinese, a YourTTS/VITS-style conditional TTS model trained at 16 kHz.
• Vocoder: the HiFiGAN waveform decoder built into the VITS checkpoint (waveform_decoder). There is no separate external vocoder checkpoint.
• Speaker conditioning: the bundled speaker encoder extracts a 512-d speaker d-vector from speaker_reference.
• Accent conditioning: walston/GenAID extracts a 64-d accent embedding from accent_reference.
• Text input: Joycent whitespace-separated pinyin phone tokens with tone digits, not Chinese characters and not full-syllable pinyin.

Main files:

• best_model.pth: AccentBox acoustic model plus built-in HiFiGAN waveform decoder.
• config.json: training config.
• speaker_encoder/model_se.pth.tar: speaker encoder checkpoint.
• speaker_encoder/config_se.json: speaker encoder config.
• remote_infer.py: batch inference from a Joycent-style jsonl list.

Input List Format

Use JSONL. Each row must contain:

{
  "id": "case_id",
  "speaker_reference": "/path/to/speaker.wav",
  "accent_reference": "/path/to/accent.wav",
  "phones": "sil n i3 h ao3 m a5 sil"
}

The Joycent inference list at resources/multi_accent/inference/medium_grl/infer_list.jsonl already has this format.

Batch Inference

Install this AccentBox codebase and its Python dependencies, then run:

python remote_infer.py \
  --repo-id walston/accentbox-zh \
  --infer-list /path/to/infer_list.jsonl \
  --output-dir outputs \
  --device cuda

For CPU smoke tests:

python remote_infer.py \
  --repo-id walston/accentbox-zh \
  --infer-list /path/to/infer_list.jsonl \
  --output-dir outputs \
  --device cpu \
  --limit 1

The script downloads this model repo, loads walston/GenAID, reads each list row, and writes one waveform per row to --output-dir.