pydrobert-kaldi

Warning

This is student-driven code, so don't expect a stable API. I'll try to use semantic versioning, but the best way to keep functionality stable is by forking.

What is it?

Some Kaldi SWIG bindings for Python. I started this project because I wanted to seamlessly incorporate Kaldi's I/O mechanism into the gamut of Python-based data science packages (e.g. Theano, Tensorflow, CNTK, PyTorch, etc.). The code base is expanding to wrap more of Kaldi's feature processing and mathematical functions, but is unlikely to include modelling or decoding.

Eventually, I plan on adding hooks for Kaldi audio features and pre-/post-processing. However, I have no plans on porting any code involving modelling or decoding.

Input/Output

Most I/O can be performed with the pydrobert.kaldi.io.open function:

>>> from pydrobert.kaldi import io
>>> with io.open('scp:foo.scp', 'bm') as f:
>>>     for matrix in f:
>>>         pass # do something

open is a factory function that determines the appropriate underlying stream to open, much like Python's built-in open. The data types we can read (here, a BaseMatrix) are listed in pydrobert.kaldi.io.enums.KaldiDataType. Big data types, like matrices and vectors, are piped into Numpy arrays. Passing an extended filename (e.g. paths to files on discs, '-' for stdin/stdout, 'gzip -c a.ark.gz |', etc.) opens a stream from which data types can be read one-by-one and in the order they were written. Alternatively, prepending the extended filename with "ark[,[option_a[,option_b...]]:" or "scp[,...]:" and specifying a data type allows one to open a Kaldi table for iterator-like sequential reading (mode='r'), dict-like random access reading (mode='r+'), or writing (mode='w'). For more information on the open function, consult the docstring. Information on Kaldi I/O can be found on their website.

The submodule pydrobert.kaldi.io.corpus contains useful wrappers around Kaldi I/O to serve up batches of data to, say, a neural network:

>>> train = ShuffledData('scp:feats.scp', 'scp:labels.scp', batch_size=10)
>>> for feat_batch, label_batch in train:
>>>     pass  # do something

Logging and CLI

By default, Kaldi error, warning, and critical messages are piped to standard error. The pydrobert.kaldi.logging submodule provides hooks into python's native logging interface: the logging module. The KaldiLogger can handle stack traces from Kaldi C++ code, and there are a variety of decorators to finagle the kaldi logging patterns to python logging patterns, or vice versa.

You'd likely want to explicitly handle logging when creating new kaldi-style commands for command line. pydrobert.kaldi.command_line provides KaldiParser, an ArgumentParser tailored to Kaldi inputs/outputs. It is used by a few command-line entry points added by this package. They are:

write-table-to-pickle: Write the contents of a kaldi table to a pickle file(s). Good for late night attempts at reaching a paper deadline.
write-pickle-to-table: Write the contents of of a pickle file(s) to a kaldi table.

Installation

If you're on a Linux, OSX, or Windows 64-bit machine and you've got Conda installed, your life is easy.

Simply:

conda install -c sdrobert pydrobert-kaldi

Which installs binaries with MKL BLAS. If nomkl is installed into the environment, OpenBLAS is used (like Numpy). Should work for Python 2.7, 3.4, 3.5, and 3.6 on Linux and OSX. Windows is limited to 3.5 and 3.6 for the time being.

Alternatively, to build through PyPI, you'll need to point the install to a BLAS library:

# for OpenBLAS
OPENBLASROOT=/path/to/openblas/install pip install \
  git+https://github.com/sdrobert/pydrobert-kaldi.git
# for MKL
MKLROOT=/path/to/mkl/install pip install \
  git+https://github.com/sdrobert/pydrobert-kaldi.git
# see setup.py for more options

You'll need either GCC or Clang plus Swig >= 3.0.8 for this.

I'd like to try to get 2.7 working for Windows, but I need c++11 support. Suggestions?

License

This code is licensed under Apache 2.0.

Code found under the src/ directory has been primarily copied from Kaldi. setup.py is also strongly influenced by Kaldi's build configuration. Kaldi is also covered by the Apache 2.0 license; its specific license file was copied into src/COPYING_Kaldi_Project to live among its fellows.

t13m / pydrobert-kaldi