YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
CA-TTS dataset packs for CSM-1b + Higgs-TTS-2 (2026-08-31)
1.04M utts / ~1,565 h (tg 543h, uz 429h, kk 585h, ky 7.6h) from existing proven corpora. Audio stays on ca-factory; this repo = manifests + code (paths are box-local).
Training contract (verified from transformers source 2026-08-31)
- BOTH models need 24 kHz MONO; processors DO NOT resample. Wavs are 16k (24k for tg_cbbook, 44.1k tg_premium) -> training script MUST cast: datasets Audio(sampling_rate=24000).
- CSM: rows -> conversation turns role=str(speaker_int); apply_chat_template(output_labels=True); model.codec_model.eval(); no shipped collator (write one, padding_side=left). Budget 2048 pos @12.5 fps audio. type=chain rows = consecutive same-narrator sentences (2-3 turns, <=45 s).
- Higgs: use_text_head=True at from_pretrained; FULL FT not LoRA (Boson guidance for new langs); type=clone rows -> ChatML [user ref_text]->[assistant ref_audio]->[user text]->[assistant audio]; 25 fps, 2048 pos budget.
- eval.jsonl per pack = 200/lang deterministic speaker-stratified holdout. Score WER with Whisper-CA (NEVER base whisper for uz — orthography trap).
- Known gotchas: bf16 NaN on padded batches (Qwen3-ASR lesson) -> smoke with real full-length batches, fp32 fallback; save_total_limit deletes WER-best ckpt -> staged eval copies.
- Downloads last month
- 21