dingke/kaldi-decoders

Custom decoders for Kaldi

kaldi-decoders

This repository contains a set of custom Kaldi decoders.

The directory has a similar structure of Kaldi's src dir: i.e. gmmbin contains decoders using Gaussian Mixture Models emissions.

Compile & install

You need to define the environment variable KALDI_ROOT to point to your Kaldi distribution.

export KALDI_ROOT=/path/to/your/kaldi/distribution
make depend
make

Once compiled, you can install the binary to PREFIX/bin (by default PREFIX=/usr/local):

make install

Available decoders

LazyLM decoders

Kaldi decoders typically recive the HCLG finite state transducer to perform the decoding. However, when a large vocabulary is used, or even more, when a large n-gram language model is used for G, the amount of memory required to store the HCLG transducer may be too large.

These decoders receive the HCL and G transducers separately, and then make a dynamic (on-the-fly) composition to obtain the HCLG transducer while decoding. The dynamic composition, together with beam pruning, makes the amount of required memory much smaller.

• decode-lazylm-faster-mapped: Decode utterances, reading log-likelihoods as matrices.
• gmm-decode-lazylm-faster: Decode features using GMM-based model.
• gmm-latgen-lazylm-faster: Generate lattices using GMM-based model.
• latgen-lazylm-faster-mapped: Generate lattices, reading log-likelihoods as matrices.