Make sure docker is installed on your computer
Clone (download) this repository on your local computer
$ git clone https://github.com/tann9949/gowajee.git
$ # to update version to latest commit,
$ # run `git pull`Register and download SRILM on the following link. Then, store it in docker directory. For this project, we use SRILM version 1.7.3.
docker
├─dockerfile
└─srilm-1.7.3.tar.gz
To download the Gowajee Corpus, check the official release repository from this link. Download it and place it to your preferred directory as we will use this dataset to mount to docker container.
We provide a dockerfile to build a Kaldi docker containing SRILM and sequitur G2P. We provide a shell script that will both build and run docker at the same time.
Copy downloaded directory which includes audios, train, dev, and test. lu set is optional as we won't use it here.
Once you finish changing the path, run run_kaldi.sh script to run Kaldi.
$ bash run_kaldi.shWe provide run.sh script that will execute the pipeline starting from preparing data up to training the model
$ bash run.shSee here for more details.
We train the model using mfcc_pitch feats on voxforge recipe. The following table shows the experiment results where LMWT were set to 17 for all models since it yields the best WER so far.
| Model | dev WER | test WER |
|---|---|---|
| mono | 40.98% | 22.71% |
| tri1 | 33.95% | 19.71% |
| tri2a | 33.76% | 19.71% |
| tri2b | 31.26% | 19.87% |
| tri2b_mmi (it3) | 32.21% | - |
| tri2b_mmi_b0.05 (it3) | 31.89% | - |
| tri2b_mpe | 31.67% | 19.83% |
| tri3b (si) | 31.42% | 17.08% |
| tri3b_mmi (si) | 31.42% | 17.08% |
| tri3b* | 21.30% | 10.52% |
| tri3b_fmmi_b (iter3)* | 19.70% | - |
| tri3b_fmmi_c (iter4)* | 20.41% | - |
| tri3b_fmmi_d (iter4)* | 20.62% | 10.27% |
| tri3b_mmi* | 22.09% | 11.36% |
| tri3b_mmi (decode2)* | 21.55% | 11.36% |
*denotes speaker dependent training
Ekapol Chuangsuwanich, Atiwong Suchato, Korrawe Karunratanakul, Burin Naowarat, Chompakorn CChaichot,and Penpicha Sangsa-nga. Gowajee Corpus. Technical report, Chulalongkorn University, Faculty of Engineering,Computer Engineering Department, 12 2020
Chompakorn Chaksangchaichot