AudioTensor

A torch.Tensor subclass that remembers the hop_length (samples per frame) of frame-based audio features, so tensors at different frame rates can be combined directly.

pip install audiotensor

import torch
from audiotensor import AudioTensor, interpolation_mode

f0 = AudioTensor(torch.randn(4, 100), hop_length=256)       # (B, T) at hop 256
loudness = AudioTensor(torch.randn(4, 400), hop_length=64)  # (B, T) at hop 64

x = f0 * loudness       # f0 is upsampled to hop 64 first
print(x.hop_length)     # 64

with interpolation_mode("cubic"):  # "linear" (default), "nearest", "cubic"
    y = f0 + loudness

• Wrap your tensor as AudioTensor(tensor, hop_length) and use it as a regular tensor.
• For ndim > 1, the shape is (B, T, ...), where B is the batch size and T is the number of frames.
• Results with ndim < 2 (e.g. x.sum(), x[0]) have no time axis and get hop_length = -1.
• Tensors with different hop_length are automatically resampled to the greatest common divisor of their hop_lengths and truncated to the shortest one when you combine them.
• In-place operations (x += y, x.add_(y), ...) resample the other operands to x, never x itself, so x must have the finest hop_length.
• hop_length always describes dim 1. Ops that move or remove that axis (transpose, reshape, ...) keep the hop_length unchanged, so use as_tensor() first if that is not what you want.
• set_hop_length(hop_length) resamples the tensor to the new hop_length.
• reduce_hop_length() upsamples the tensor to hop_length = 1.
• steps returns the size of T in the tensor.
• truncate(steps) truncates the tensor to the given steps.
• unfold(size, step) returns a view of the tensor with a sliding window of size size and step step along the T dimension.
• as_tensor() returns a plain torch.Tensor.