stopamertime/pitchscape-models

Pitchscape's models

The trained models Pitchscape runs, in the browser (ONNX Runtime Web on WebGPU) and on its server, converted to ONNX. Each keeps its own authors' licence:

file	what it is	licence	size
silero_vad_v6.onnx	Silero VAD v6 voice activity detector (Silero Team, https://github.com/snakers4/silero-vad): the ONNX file faster-whisper 1.2.1 ships and runs for vad_filter, copied unchanged.	MIT	1.2 MB
whisper_medium_encoder_f32.onnx	OpenAI Whisper medium (A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, I. Sutskever, "Robust Speech Recognition via Large-Scale Weak Supervision", 2022), the weights of https://huggingface.co/openai/whisper-medium, its encoder exported to ONNX in float32 (for ONNX Runtime's CPU provider) by tools/export_whisper.py.	MIT	1430.4 MB
whisper_medium_decoder_f32.onnx	OpenAI Whisper medium (A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, I. Sutskever, "Robust Speech Recognition via Large-Scale Weak Supervision", 2022), the weights of https://huggingface.co/openai/whisper-medium, its decoder exported to ONNX in float32 by tools/export_whisper.py.	MIT	1625.3 MB
whisper_medium_encoder.onnx	OpenAI Whisper medium (A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, I. Sutskever, "Robust Speech Recognition via Large-Scale Weak Supervision", 2022), the weights of https://huggingface.co/openai/whisper-medium, its encoder (and each decoder layer's cross-attention keys and values) exported to ONNX in float16 by tools/export_whisper.py.	MIT	715.3 MB
whisper_medium_decoder.onnx	OpenAI Whisper medium (A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, I. Sutskever, "Robust Speech Recognition via Large-Scale Weak Supervision", 2022), the weights of https://huggingface.co/openai/whisper-medium, its decoder exported to ONNX in float16 by tools/export_whisper.py.	MIT	812.7 MB
whisper_medium.json	OpenAI Whisper medium (A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, I. Sutskever, "Robust Speech Recognition via Large-Scale Weak Supervision", 2022), the weights of https://huggingface.co/openai/whisper-medium, its tokenizer and the tokens faster-whisper suppresses, as JSON (from https://huggingface.co/Systran/faster-whisper-medium).	MIT	0.8 MB
htdemucs_6s.onnx	Demucs htdemucs_6s (Hybrid Transformer Demucs, 6 sources), Meta AI, from https://github.com/facebookresearch/demucs (demucs 4.1.0). Exported to ONNX without its STFT and iSTFT by tools/export_models.py.	MIT	114.6 MB
adtof_frame_rnn.onnx	ADTOF Frame_RNN drum transcription (M. Zehren, M. Alunno, P. Bientinesi, "High-Quality and Reproducible Automatic Drum Transcription from Crowdsourced Data", Signals 2023), checkpoint Frame_RNN_adtofAll_0 from https://github.com/MZehren/ADTOF, as converted to ONNX (with madmom's spectrogram in the graph) at https://huggingface.co/thebluescreen/adtof-drums-onnx; copied unchanged. Non-commercial use only.	CC BY-NC-SA 4.0	2.2 MB
crepe_full.onnx	CREPE pitch tracker, model "full" (J. W. Kim, J. Salamon, P. Li, J. P. Bello, "CREPE: A Convolutional Representation for Pitch Estimation", ICASSP 2018), with the weights torchcrepe 0.0.24 ships (https://github.com/maxrmorrison/torchcrepe). Exported to ONNX (frames in, the network's 360 pitch-bin probabilities out) by tools/export_models.py.	MIT	89.0 MB
basic_pitch.onnx	Basic Pitch ICASSP 2022 polyphonic note transcription (R. Bittner, J. J. Bosch, D. Rubinstein, G. Meseguer-Brocal, S. Ewert, "A Lightweight Instrument-Agnostic Model for Polyphonic Note Transcription and Multipitch Estimation", ICASSP 2022), Spotify, from https://github.com/spotify/basic-pitch: the ONNX file the basic-pitch 0.4.0 package ships, copied unchanged.	Apache-2.0	0.2 MB
align_en.onnx	wav2vec 2.0 base fine-tuned for English on LibriSpeech 960 h (A. Baevski, H. Zhou, A. Mohamed, M. Auli, "wav2vec 2.0", 2020; fairseq), as torchaudio 2.7.1 ships it (pipelines.WAV2VEC2_ASR_BASE_960H), exported to ONNX in float16 with log_softmax by tools/export_align.py.	MIT	189.0 MB
align_en_f32.onnx	wav2vec 2.0 base fine-tuned for English on LibriSpeech 960 h (A. Baevski, H. Zhou, A. Mohamed, M. Auli, "wav2vec 2.0", 2020; fairseq), as torchaudio 2.7.1 ships it (pipelines.WAV2VEC2_ASR_BASE_960H), exported to ONNX in float32 (for ONNX Runtime's CPU provider) with log_softmax by tools/export_align.py.	MIT	377.8 MB
align_mms.onnx	the MMS forced aligner (V. Pratap et al., "Scaling Speech Technology to 1,000+ Languages", 2023), as torchaudio 2.7.1 ships it (pipelines.MMS_FA), exported to ONNX in float16 with log_softmax by tools/export_align.py.	CC BY-NC 4.0	631.4 MB
align_mms_f32.onnx	the MMS forced aligner (V. Pratap et al., "Scaling Speech Technology to 1,000+ Languages", 2023), as torchaudio 2.7.1 ships it (pipelines.MMS_FA), exported to ONNX in float32 (for ONNX Runtime's CPU provider) with log_softmax by tools/export_align.py.	CC BY-NC 4.0	1262.3 MB

manifest.json lists each file with its size and SHA-256. Made by tools/export_models.py in that repository.