This repository contains Python helper functions to read/write feature and label files for machine learning.

Supported file types include HTK feature files, Janus feature files, Kaldi archive and script files, pfiles, and TextGrid annotation files.

Below is a brief documentation of the functions:

• file = smart_open(filename, mode, ...)

  Works just like the built-in open function, but can deal with .gz and .bz2 compressed files transparently.

• x = readHtk(filename)

  Reads the feature matrix stored in a HTK feature file, and returns it as a 2-D numpy array.

• writeHtk(filename, feature, sampPeriod, parmKind):

  Writes a 2-D numpy array to a HTK feature file. sampPeriod and parmKind are required fields of HTK feature files.

• x = readFmatrix(filename)

  Reads the feature matrix stored in a Janus feature file, and returns it as a 2-D numpy array.

• writeFmatrix(filename, feature)

  Writes a 2-D numpy array to a Janus feature file.

• features, uttids = readArk(filename, limit = numpy.inf)

  Reads at most limit feature matrices in a Kaldi ark file.
  Returns the feature matrices in a list of 2-D numpy arrays, and their utterance ids in a list of strings.

• features, uttids = readScp(filename, limit = numpy.inf)

  Reads at most limit feature matrices in a Kaldi scp file (which contains pointers into ark files).
  Returns the feature matrices in a list of 2-D numpy arrays, and their utterance ids in a list of strings.

• writeArk(filename, features, uttids)

  Takes a list of 2-D numpy arrays and a list of utterance ids, and writes them into a Kaldi ark file.
  Returns a list of strings in the format "filename:offset", which can be used to write a Kaldi script file.

• writeScp(filename, uttids, pointers)

  Takes a list of utterance ids and a list of pointer strings in the format "filename:offset", and writes them to a Kaldi scp file.

• features, labels = readPfile(filename)

  Reads the feature matrices (and labels) stored in a pfile.
  Returns both the features and labels as a list of 2-D numpy arrays (although in most cases the label matrices will have a width of 1).
  In the case where the pfile doesn't contain labels, labels will be None.

• writePfile(filename, features, labels = None)

  Writes feature matrices (and labels) to a pfile.
  Both the features and labels should be presented as lists of 2-D numpy arrays.
  In the case where there is only one label per frame, the label arrays can be 1-D.

• annot = TextGrid.fromFile(filename)

  Reads the annotations in a TextGrid file, and returns it as an annotation object.
  An annotation object consists of tiers, which in turn consists of intervals, which have starting times, ending times, and labels.