GoParrot is a simple tool that computes GOP (Goodness of Pronunciation) scores based on kaldi for oral reading assessment. It provides the baseline for Interspeech paper "ASR-Free Pronunciation Assessment".
Name it parrot as parrot repeats what you say (鹦鹉学舌).
Three kinds of gops can be computed, based on posterior, likelihood and likelihood ratio separately.
Posterior
After force-alignment, the log posterior of each phone is the average log posterior of all frames belong to that phone. The GOP score is the average log posterior of all phones. See eq.(7) in this paper.
Likelihood
After force-alignment, the log likelihood of each phone is the average log likelihood of all frames belong to that phone. The GOP score is the average log likelihood of all phones. See eq.(18) in this paper.
Likelihood ratio
After force-alignment, the likelihood of each frame can be scaled by dividing the max likelihood of each frame, then the log likelihood ratio of each phone is the average log likelihood ratio of all frames belong to that phone. The GOP score is the average log likelihood ratio of all phones. See eq.(16) involving eq.(18) and (19) in this paper.
A Kaldi ASR trained with WSJ dataset is provided, and two samples are given.
git clone https://github.com/tzyll/goparrot.git
cd goparrot
# compile kaldi in the dir, or give a softlink 'kaldi' to the dir
# you may just uncomment the following lines to clone a kaldi
# then follow its instruction to install
# git submodule init
# git submodule update
pip3 install kaldi_io
./run.shJust prepare test data with kaldi style as in data/test which includes at least
wav.scp, text, utt2spk and spk2utt (utt2spk and spk2utt can be fake, i.e., just wav-id to wav-id).
Train ASR system with Kaldi with your own dataset, and replace data/lang and exp/nnet3/tdnn with your own.
We scored hundreds of utterances from ERJ, and the Pearson correlation coefficient (PCC) between human scores and the three types of GOP scores were 0.6249, 0.6334 and 0.6361 separately. The PCC between human scores were about 0.57.
The given ASR model was trained with cross-entropy loss. We also scored with Kaldi Chain (LF-MMI) model, and the PCC between human scores and the three types of GOP scores were 0.1615, 0.1615 (chain model has no prior, so likelihood is actually posterior) and 0.0934 separately.