This repository provides the recipe for the paper A Crowdsourced Open-Source Kazakh Speech Corpus and Initial Speech Recognition Baseline.
Our code builds upon ESPnet, and requires prior installation of the framework. Please follow the installation guide and put the ksc folder inside espnet/egs/ directory.
After succesfull installation of ESPnet & Kaldi, go to ISSAI_SAIDA_Kazakh_ASR/asr1 folder and create links to the dependencies:
ln -s ../../../tools/kaldi/egs/wsj/s5/steps steps
ln -s ../../../tools/kaldi/egs/wsj/s5/utils utils
The directory for running the experiments (ISSAI_SAIDA_Kazakh_ASR/<exp-name) can be created by running the following script:
./setup_experiment.sh <exp-name>
Download ISSAI_KSC_335RS dataset and untar in the directory of your choice. Specify the path to the dataset inside ISSAI_SAIDA_Kazakh_ASR/<exp-name>/conf/data_path.conf file:
dataset_dir=/path-to/ISSAI_KSC_335RS_v1.1
To train the models, run the script ./run.sh inside ISSAI_SAIDA_Kazakh_ASR/<exp-name>/ folder.
You can find the link to the latest pre-trained Transformer model here. Untar it in ksc/<exp-name>/.
To decode a single audio, specify paths to the following files inside recog_wav.sh script:
lang_model= path to rnnlm.model.best
cmvn= path to cmvn.ark for example data/train/cmvn.ark
recog_model= path to e2e model, in case of transformer: model.last10.avg.best
Then, run the following script:
./recog_wav.sh <path-to-audio-file>