"""Torchaudio sinc kernels with integer-exact output length for long files.""" from torch.nn import functional as F def output_length(samples, source_rate, target_rate): return (int(samples) * int(target_rate) + int(source_rate) - 1) // int(source_rate) def resample_exact(waveform, transform): if transform.orig_freq == transform.new_freq: return waveform original = transform.orig_freq // transform.gcd new = transform.new_freq // transform.gcd shape = waveform.shape packed = waveform.reshape(-1, shape[-1]) padded = F.pad(packed, (transform.width, transform.width + original)) result = F.conv1d(padded[:, None], transform.kernel, stride=original) result = result.transpose(1, 2).reshape(packed.shape[0], -1) target = output_length(shape[-1], original, new) assert result.shape[-1] >= target return result[..., :target].reshape(shape[:-1] + (target,))