respiro-en-onnx

• Original work: Respiro-en by DongYANG (GitHub: ydqmkkx)
• License: MIT (see LICENSE, preserving the original copyright)
• This repo only redistributes an ONNX export of the original respiro-en.pt checkpoint. No retraining or architectural change was performed; the network is numerically identical to the original within float tolerance (max abs diff of the per-frame probabilities vs. the original PyTorch model ≈ 3e-6 across the upstream sample clips).

Model

• File: respiro-en.onnx (opset 17)
• Input: feat — float32 [1, 3, 128, T]
  • channel 0: log-mel spectrogram in dB (librosa.power_to_db(melspectrogram(...), ref=np.max))
  • channel 1: variance of the log-mel across the frequency axis (unbiased), broadcast over mels
  • channel 2: zero-crossing rate, broadcast over mels
  • feature params: 16 kHz, n_fft=400, hop_length=160 (10 ms hop), n_mels=128, Slaney mel scale (htk=False), power=2.0, power_to_db(ref=np.max, top_db=80)
• Output: prob — float32 [1, T], per-frame breath probability in [0, 1] (10 ms hop)
• Dynamic axis: T (sequence length) is dynamic.

Note on the export: the original network feeds a length tensor to a torchaudio Conformer purely to build a self-attention padding mask. For a single, full-length (batch=1, unpadded) input that mask is all-valid, so it has no effect on the output. To make the exported graph robust to arbitrary T, the torch.nn.MultiheadAttention modules were replaced with an equivalent dynamic-shape implementation (same weights) before export, which removed the need for the length input. The exported model therefore has a single input, feat.

Conversion

Attribution