ggaps/sound-space-models

Sound Space – browser models

ONNX files used by the Sound Space web app to analyse music entirely inside the user's browser (onnxruntime-web, WebGPU). Audio never leaves the user's device; only these model files are downloaded.

Large weights are stored as float16 and cast back to float32 at load time (Cast nodes), so computation is float32.

File	What	Original project	License
htdemucs_6s.onnx	Demucs v4 6-stem separation core, STFT/iSTFT moved outside the graph (sevagh/demucs.onnx method)	facebookresearch/demucs (htdemucs_6s)	MIT, Copyright (c) Meta Platforms, Inc. and affiliates
htdemucs_6s_mobile.onnx	Same htdemucs_6s weights exported with a 2.93 s segment (129024 samples) for phones: CPU (WebAssembly) peak memory during a run drops from about +1.7 GB to about +190 MB	facebookresearch/demucs (htdemucs_6s)	MIT, Copyright (c) Meta Platforms, Inc. and affiliates
basic_pitch_nmp.onnx	Basic Pitch note/onset model (unchanged)	spotify/basic-pitch	Apache-2.0, Copyright Spotify AB
beat_this_final0.onnx, beat_this_melfb.f32	Beat This! beat/downbeat tracker (final0) and its mel filterbank	CPJKU/beat_this	MIT (code and published weights)
drumsep.onnx	Drum stem split into kick / snare / cymbals / toms (HDemucs), STFT outside the graph	inagoy/drumsep	Repository MIT. The weights are distributed by the author via Google Drive without a separate license statement.
clap_audio.onnx	CLAP audio encoder + projection (input: log-mel [1,1,1001,64] at 48 kHz, output: L2-normalised 512-d embedding). Mood text embeddings are precomputed in the app	laion/clap-htsat-unfused (Hugging Face)	Apache-2.0
onset_mel_44k.f32	librosa mel filterbank (44.1 kHz, n_fft 2048, 128 bands)	librosa	ISC

Conversion scripts and verification (ONNX vs PyTorch differences, evaluation results) are in the Sound Space project (scripts/export_demucs_onnx.sh, scripts/export_beat_this_onnx.py, scripts/export_drumsep_onnx.py, scripts/export_clap_onnx.py, scripts/onnx_fp16_weights.py).