Python Speech Features

0.6 · active · verified Sat Apr 11

python-speech-features is a Python library designed for extracting common speech features used in Automatic Speech Recognition (ASR). It provides functionalities to compute Mel-Frequency Cepstral Coefficients (MFCCs), filterbank energies, log filterbank energies, and spectral subband centroids. The current stable version on PyPI is 0.6, last released in 2017, with a slightly newer v0.6.1 tag on its GitHub repository from 2020. The project maintains a slow release cadence, but its core functionalities remain widely used for fundamental speech feature extraction.

Warnings

gotcha The PyPI version (0.6, last updated Aug 2017) is older than the latest tag on GitHub (v0.6.1, Jan 2020). Users installing via `pip install python-speech-features` might not get the absolute latest code, which could have minor fixes or changes not yet reflected on PyPI.
Fix: For the very latest version, consider installing directly from the GitHub repository: `pip install git+https://github.com/jameslyons/python_speech_features.git`.
gotcha When integrating with other audio processing libraries like `librosa`, be aware of data type expectations. `scipy.io.wavfile.read` typically returns `int16` samples, while some libraries might expect `float32`. Additionally, MFCC computation methodologies can differ between libraries (e.g., `python-speech-features` uses Discrete Fourier Transform, `librosa` uses Short-Time Fourier Transform), leading to different output shapes or values for seemingly identical parameters.
Fix: Explicitly convert signal data types (e.g., `signal.astype(np.float32)`) if required by other libraries. Consult the documentation of both `python-speech-features` and any other library being used for parameter details and expected outputs. Be mindful that `numcep` is effectively capped by `nfilt` in `python-speech-features`.
gotcha A common warning, `WARNING:root:frame length (X) is greater than FFT size`, can occur if the product of `winlen` (window length in seconds) and `samplerate` (frame length in samples) exceeds `nfft` (FFT size).
Fix: Ensure that `nfft` is greater than or equal to `winlen * samplerate`. For example, if `winlen=0.025` and `samplerate=16000`, the frame length is 400 samples. `nfft` should be set to 512, 1024, or a higher power of 2 that accommodates this length.

Install

pip install python-speech-features Install from PyPI

Imports

mfcc
```
from python_speech_features import mfcc
```
fbank
```
from python_speech_features import fbank
```
Used to compute Mel-filterbank energy features.
logfbank
```
from python_speech_features import logfbank
```
Used to compute log Mel-filterbank energy features.
ssc
```
from python_speech_features import ssc
```
Used to compute Spectral Subband Centroid features.

Quickstart

This quickstart demonstrates how to generate a simple audio signal, save it as a WAV file, and then use `python-speech-features` to extract both Mel-Frequency Cepstral Coefficients (MFCCs) and log Mel-filterbank energies. It uses `scipy.io.wavfile` to handle audio file I/O.

import numpy as np
from scipy.io import wavfile
from python_speech_features import mfcc, logfbank
import os

# Create a dummy WAV file for demonstration
samplerate = 16000 # Hz
duration = 1 # seconds
f_hz = 440 # A4 note

t = np.linspace(0., duration, int(samplerate * duration))
signal = 0.5 * np.sin(2 * np.pi * f_hz * t)

# Scale to 16-bit integer for WAV file
wav_signal = (signal * 32767).astype(np.int16)
dummy_wav_filename = 'dummy_audio.wav'
wavfile.write(dummy_wav_filename, samplerate, wav_signal)

# Read the audio file
(rate, sig) = wavfile.read(dummy_wav_filename)

# Compute MFCC features
mfcc_feat = mfcc(sig, rate)
print(f"MFCC features shape: {mfcc_feat.shape}")

# Compute Log Filterbank energies
fbank_feat = logfbank(sig, rate)
print(f"Log Filterbank features shape: {fbank_feat.shape}")

# Clean up the dummy file
os.remove(dummy_wav_filename)

view raw JSON →