TTS data cleaning compared: how Fish Audio, Emilia, CosyVoice 3 and NeMo Curator pipelines differ
solyarisoftware · x · 2026-10-26
@JinyuChai shared a comparison infographic of four TTS data-cleaning pipelines, arguing that where your audio comes from shapes how you clean it: background music, multilingual speech, and uneven recording quality each demand different checks beyond the usual VAD, diarization, and ASR.
- Fish Audio S2: vocal/music separation → VAD splitting → custom SQM quality gate → rich transcription with speaker turns and vocal tags (<laugh>, <inhale>).
- Emilia-Pipe: UV vocal separation → pyannote + Silero VAD → Whisper transcription → DNSMOS quality and language-confidence filtering → character/phone duration proportion checks.
- CosyVoice 3: segmentation + denoising with boundary checks → cross-transcription with Whisper, Canary, and SeamlessM4T to filter disagreements → MFA punctuation alignment → volume and token-length ratio filters.
- NeMo Curator: built for large mixed collections, with normalize + VAD → UTMOOS/SIGMOS quality pass → per-speaker grouping and rechecking → timestamp-preserving clip export plus ASR + text checks.
Takeaway: each pipeline allocates effort differently across audio cleaning, quality checks, text/timing alignment, and per-speaker checks depending on its source data.
More from Multimodal
- Real-time fluid sim demo ships 9 looks and 11 physics presets with GPU-baked collision — Michael_Moroz_ · 2026-10-26
- AI-generated frog romance short with a bee punchline delights Reddit — HarshaShastry · 2026-10-26
- Adobe's SliderEdit++ Improves Slider-Based Image Editing by 18.5% with New SlideBench — Arman Zarei · 2026-10-26
- 'La dolce vita': egeberkina shares retro Italian-summer AI image set — egeberkina · 2026-10-26
- AI short film teaser 'Kaalakund' shows near-photoreal moody landscapes — KalpnkAIStudio · 2026-10-26
- AI art workflow: The Streets made with Midjourney, GPT Image2 and Seedance 2.5 — Salmaaboukarr · 2026-10-26