timur777/singcheck-models

singcheck models

The networks the singcheck app runs on the user's own computer to prepare a song for singing along: the singer's voice alone, her melody, the words with their times, the exact word starts, who sings each line, the beats. ONNX exports (made with scripts/export_onnx.py of the singcheck project, each checked against the original PyTorch model); the app fetches them once (about 2.4 GB) and runs them with ONNX Runtime — no audio leaves the computer. Big weights are stored in float16 and turned into float32 when loaded (half the download, the same results).

File	What	Source	License
htdemucs.onnx	the voice alone: htdemucs between its STFT and iSTFT (its 1D convolutions as 2D)	facebookresearch/demucs	MIT
crepe_full.onnx	the singer's melody: CREPE full	marl/crepe via torchcrepe	MIT
parakeet-fp16w/	words with timestamps: Parakeet TDT 0.6B v3; nemo128.onnx is onnx-asr's preprocessor	nvidia/parakeet-tdt-0.6b-v3 via istupakov/parakeet-tdt-0.6b-v3-onnx, onnx-asr	CC-BY-4.0 (model), MIT (onnx-asr)
mms_fa.fp16w.onnx (+ .data)	the lyrics aligner: MMS_FA (wav2vec2, with the star token)	torchaudio MMS_FA, Pratap et al. 2023, Scaling Speech Technology to 1,000+ Languages	CC-BY-NC-4.0 — non-commercial use only
ecapa.onnx	a speaker print (with its Fbank): ECAPA-TDNN	speechbrain/spkrec-ecapa-voxceleb	Apache-2.0
beat_this.onnx	the beats: Beat This! (final0) on 1500-frame chunks	CPJKU/beat_this	MIT

Each file keeps its upstream license; with MMS_FA in it the set as a whole is for non-commercial use.