FCPE (ONNX)
An ONNX export of the pitch estimator FCPE by CN_ChiTu: the model bundled with
torchfcpe 0.0.4 (fcpe_c_v001.pt, SHA-256
b9aeaeb673436eeda50ceafd632aa681aa63417e52eae4207503d180c9b10015), unchanged. Exported for Tune Studio.
fcpe.onnx: input audio [1, samples] float32, 16 kHz mono. Output f0 [1, samples // 160 + 1], the pitch
in Hz every 10 ms, 0 where unvoiced. The mel spectrogram and the decoding are inside the graph, with
decoder_mode="local_argmax", threshold=0.006 and a ceiling of 1400 Hz. The spectrogram is computed with plain
convolutions rather than the STFT op, so every execution provider can run it.
Checked against torchfcpe on CPU on two full songs: voiced/unvoiced identical on every frame, and one frame in 17,953 differs by more than 1 cent (11 cents).
License
MIT, like the original (see LICENSE).