Torgo ASR

Description

This is a Kaldi recipe to build automatic speech recognition systems on the Torgo corpus of dysarthric speech.

Setup

Update the KALDI_ROOT and DATA_ORIG variables in path.sh to point to the correct locations for your Kaldi installation and the Torgo corpus. Then run the following:

source path.sh
ln -s $KALDI_ROOT/egs/wsj/s5/{steps,utils} .

Some scripts in local/ also require the following Python packages:

invoke numpy pandas python-Levenshtein

Usage

The following instructions allow to train ASR systems on Torgo and to reproduce results from the paper.

Train ASR systems

# HMM/GMM systems:
./run.sh

# LF-MMI (TDNN-F) systems:
./run_tdnnf.sh

# CE (TDNN-LSTM) systems:
./local/nnet3/run_tdnn_lstm.sh

# Show WER:
./local/get_wer.py exp/sgmm

Corpus statistics

Torgo corpus statistics:

./local/corpus_statistics.sh

Pronunciation similarity

How similar are the isolated words to each other? First retrieve the phonetic representation for each word, then analyse the similarity of pronunciations:

./local/get_prons.sh > data/pronunciations_single
./local/compute_pron_similarity.py

Phone duration

We analysed how mean phoneme duration and WER are correlated.

# Get phone alignments with duration information:
./local/get_phone_alignments.sh exp/sgmm

# Compute mean phoneme durations:
./local/analyze_phone_lengths.py

Citation

Please cite the following paper if you use this code for your research.

@inproceedings{hermann2020.asr,
    author = "Hermann, Enno and Magimai.-Doss, Mathew",
    title = "Dysarthric Speech Recognition with Lattice-Free {MMI}",
    booktitle = "Proceedings International Conference on Acoustics, Speech, and Signal Processing (ICASSP)",
    pages = "6109--6113",
    year = "2020",
    doi = "10.1109/ICASSP40776.2020.9053549"
}

The code is based on an earlier recipe by Cristina España-Bonet and José A. R. Fonollosa.