| # WTForbes/tw_tts_set | |
| ## Upload Status | |
| - repo_created: yes | |
| - gated_access: user will configure in Hugging Face UI | |
| - corpus_jsonl: generated by `prepare_zh_hf_upload.py` | |
| - needs_audio_upload: true | |
| ## Method | |
| - Source rows are selected only from zh-CN/zh-TW JSONL files. | |
| - Audio is uploaded only if its exact path appears in the selected JSONL rows. | |
| - The source audio directory is never uploaded wholesale. | |
| - Shared zh-CN/zh-TW repos use a union of referenced audio paths. | |
| - The same referenced audio path maps to the same deterministic audio_id. | |
| - WAV input must be converted to FLAC before being written into Parquet. | |
| - Existing FLAC/MP3/Opus/Ogg keep their codec unless changed later. | |
| ## Directory Layout | |
| ```text | |
| corpus_dir_jsonl/ | |
| train.jsonl | |
| train_stage1.jsonl | |
| eval.jsonl | |
| manifests/ | |
| audio_manifest.tsv | |
| set_summary.tsv | |
| upload_progress.json | |
| ``` | |
| ## Notes | |
| Shared zh-CN/zh-TW audio union for tw_tts_set. | |
| Known external pending item: `WTForbes/unc_5` is not present on this machine and is intentionally skipped. | |