This project computes GOP (Goodness of Pronunciation) bases on Kaldi.
- Download and complile Kaldi. Note that you need to check out the branch 5.1 instead of master, and do not use the "--shared" option.
- Edit src/CMakeLists.txt to set the variable $KALDI_ROOT.
- Compile the binary:
cd src/
mkdir build && cd build
cmake .. && make
cd egs/gop-compute
./run.sh
You can find the GOP raw scores in eval_gop/gop.1
and the corresponding phoneme sequence (in numbers) in eval_gop/phonemes.1
- Download the sample data and extract all files under
egs/gop-compute
. This package provides a toy speech dataset as well as a pretrained acousitc model - Run
./run_tri6b.sh data/sample_raw_data/
for a demo - Run
./run_tri6b.sh
for usage information - If you want to use your own data, you can organize your files similar to
data/sample_raw_data/
in the sample data package - The code assumes that the name of the root dir of your own data is the same as the speaker's ID, see
data/sample_raw_data/spk2gender
- This script also outputs the symbolic phoneme sequence in
eval_gop_tri6b/phonemes_sym.*
Huge thanks to Junbo Zhang for his initial kaldi-gop repository. I added the phoneme output and added an example on how to use another acoustic model for the computation.