spectra-torch

Spectra Extraction based on PyTorch


Keywords
filterbank, mfcc, pytorch, voice-activity-detection
Install
pip install spectra-torch==0.4.0

Documentation

spectra_torch

Considering the pytorch-kalda 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.

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)

Reference

Thanks for you attention.

Free for question to my email (rwang@tongji.edu.cn).