Author: Prem Seetharaman
An STFT/iSTFT written up in PyTorch using 1D Convolutions. Requirements are a recent version PyTorch, numpy, and librosa (for loading audio in test_stft.py). Thanks to Shrikant Venkataramani for sharing code this was based off of and Rafael Valle for catching bugs and adding the proper windowing logic. Uses Python 3.
Install easily with pip:
pip install torch-stft
import torch
from torch_stft import STFT
import numpy as np
import librosa
import matplotlib.pyplot as plt
audio = librosa.load(librosa.util.example_audio_file(), duration=10.0, offset=30)[0]
device = 'cpu'
filter_length = 1024
hop_length = 256
win_length = 1024 # doesn't need to be specified. if not specified, it's the same as filter_length
window = 'hann'
audio = torch.FloatTensor(audio)
audio = audio.unsqueeze(0)
audio = audio.to(device)
stft = STFT(
filter_length=filter_length,
hop_length=hop_length,
win_length=win_length,
window=window
).to(device)
magnitude, phase = stft.transform(audio)
output = stft.inverse(magnitude, phase)
output = output.cpu().data.numpy()[..., :]
audio = audio.cpu().data.numpy()[..., :]
print(np.mean((output - audio) ** 2)) # on order of 1e-16
Output of compare_stft.py:
Test it by just cloning this repo and running
pip install -r requirements.txt
python -m pytest .
Unfortunately, since it's implemented with 1D Convolutions, some filter_length/hop_length combinations can result in out of memory errors on your GPU when run on sufficiently large input.
Pull requests welcome.