Introduction

kaldifeat implements speech feature extraction algorithms compatible with Kaldi using PyTorch, supporting CUDA as well as autograd.

kaldifeat has the following features:

• Fully compatible with Kaldi

  Note

  The underlying C++ code is copied & modified from Kaldi directly. It is rewritten with PyTorch C++ APIs.

• Provide not only C++ APIs but also Python APIs

  Note

  You can access kaldifeat from Python.

• Support autograd

• Support CUDA and CPU

  Note

  You can use CUDA for feature extraction.

• Support online (i.e., streaming) and offline (i.e., non-streaming) feature extraction

• Support chunk-based processing

  Note

  This is especially usefull if you want to process audios of several hours long, which may cause OOM if you send them for computation at once. With chunk-based processing, you can process audios of arbirtray length.

• Support batch processing

  Note

  With kaldifeat you can extract features for a batch of audios

Currently implemented speech features and their counterparts in Kaldi are listed in the following table.

Table 1 Supported speech features

Supported speech features	Counterpart in Kaldi
kaldifeat.Fbank	compute-fbank-feats
kaldifeat.Mfcc	compute-mfcc-feats
kaldifeat.Plp	compute-plp-feats
kaldifeat.Spectrogram	compute-spectrogram-feats
kaldifeat.OnlineFbank	kaldi::OnlineFbank
kaldifeat.OnlineMfcc	kaldi::OnlineMfcc
kaldifeat.OnlinePlp	kaldi::OnlinePlp

Each feature computer needs an option. The following table lists the options for each computer and the corresponding options in Kaldi.

Hint

Note that we reuse the parameter names from Kaldi.

Also, both online feature computers and offline feature computers share the same option.

Table 2 Feature computer options

Options in kaldifeat	Corresponding options in Kaldi
kaldifeat.FbankOptions	kaldi::FbankOptions
kaldifeat.MfccOptions	kaldi::MfccOptions
kaldifeat.PlpOptions	kaldi::PlpOptions
kaldifeat.SpectrogramOptions	kaldi::SpectrogramOptions
kaldifeat.FrameExtractionOptions	kaldi::FrameExtractionOptions
kaldifeat.MelBanksOptions	kaldi::MelBanksOptions

Read more to learn how to install kaldifeat and how to use each feature computer.