Skip to content

mechanicalsea/spectra

Repository files navigation

spectra_torch

Considering the pytorch-kaldi is presented, so it is more practical to use it. Also, SpeechBrain, A PyTorch-based Speech Toolkit, is coming. I am looking forward to a nice step on speech. To conclude, this package is used to learn spectra of a signal, so it is valuable at all.

News: Tutorials continue to come! Jupiter Notebook Viewer for "Reaload?"er.

This library provides common spectra features from an audio signal including MFCCs and filter bank energies. This library mimics the library python_speech_features but PyTorch-style.

This library provides voice activity detection (VAD) based on energy. This library mimics the library VAD-python but PyTorch-style.

Use: Rui Wang. (2020, March 14). mechanicalsea/spectra: release v0.4.0 (Version 0.4.0).

Installation

This library is avaliable on pypi.org

To install from Pypi:

pip install --upgrade spectra-torch

Require:

  • python: 3.7.3
  • torch: 1.4.0
  • torchaudio: 0.4.0

Usage

Supported features:

  • Mel Frequency Cepstral Coefficients (MFCC)
  • Filterbank Energies
  • Log Filterbank Energies
  • Voice Activity Detection (VAD)

Here are examples.

Easy demo:

# Ensure cuda is available.
import spectra_torch.base as mm
import torchaudio as ta

sig, sr = ta.load_wav('piece_20_32k.wav')
sig = sig[0].cuda()
mfcc = mm.mfcc(sig, sr) # MFCC
starts, detection = mm.is_speech(sig, sr, speechlen=0.5) # VAD

Tutorial

Tutorials of MFCC and VAD is provided at notebooks.

Step-by-step description is presented. Welcome to enjoy it.

Performance

The difference between spectra_torch and python_speech_features:

  • Precision bais: 1e-4
  • Speed up: 0.1s/mfcc

MFCC

def mfcc(signal, samplerate=16000, winlen=0.025, hoplen=0.01, 
         numcep=13, nfilt=26, nfft=None, lowfreq=0, highfreq=None, 
         preemph=0.97, ceplifter=22, plusEnergy=True)

Filterbank

def fbank(signal, samplerate=16000, winlen=0.025, hoplen=0.01, 
          nfilt=26, nfft=512, lowfreq=0, highfreq=None, preemph=0.97)

VAD

def is_speech(signal, samplerate=16000, winlen=0.02, hoplen=0.01, 
              thresEnergy=0.6, speechlen=0.5, lowfreq=300, highfreq=3000, 
              preemph=0.97)

Parameterized Bandpass Filter

class PFilter(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, stride=1, 
                 padding=0, dilation=1, groups=1, bias=False, device="cpu",
                 mode='bandpass',sample_rate=16000, min_hz=50, max_hz=None,
                 min_band_hz=50, win_fn="Hamming")

Reference

Thanks for you attention.

Free for question to my email ([email protected]).

About

Spectra extraction tutorials based on torch and torchaudio.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published