pydrobert-speech 0.0.1

Speech processing with Python

pydrobert-speech

This pure-python library allows for flexible computation of speech features.

For example, given feature configuration called fbanks.json:

{
  "name": "stft",
  "bank": "fbank",
  "frame_length_ms": 25,
  "include_energy": true,
  "pad_to_nearest_power_of_two": true,
  "window_function": "hanning",
  "use_power": true
}

You can compute triangular, overlapping filters like Kaldi or HTK with the commands

import json
from pydrobert.speech import *
# get the feature computer ready
params = json.load(open('fbank.json'))
computer = util.alias_factory_subclass_from_arg(compute.FrameComputer, params)
# assume "signal" is a numpy float array
feats = computer.compute_full(signal)

If you plan on using a PyTorch DataLoader or Kaldi tables in your ASR pipeline, you can compute all a corpus' features by using the commmands signals-to-torch-feat-dir (requires pytorch package) or compute-feats-from-kaldi-tables (requires pydrobert-kaldi package).

This package can compute much more than f-banks, with many different permutations. Consult the documentation for a more in-depth discussion of how to use it.

Documentation

• Latest
• Stable

Installation

pydrobert-speech is available via both PyPI and Conda.

conda install -c sdrobert pydrobert-speech
pip install pydrobert-speech
pip install git+https://github.com/sdrobert/pydrobert-speech # bleeding edge