junhokim94 / asr_project

This repository created for the NHN ASR hackathon competition.

ASR_project

This repository created for ASR Hackathon Competition.

We won the Excellence award in NHN Consortium 2020 AI Training Data Hackathon Competition (코로나음성인식 Team) http://hackathon.workpedia.co.kr/

**We customized espnet1, espnet2 https://github.com/espnet/espnet

Dataset information

Training data information is as follows NHN ASR hackerthon dataset (not publicly available now) Voice data types consist of men and women(adult, senior, children, foreign languages). It is free-talk voice data(total of 200 speakers and approximately 400 hours of data). The PCM data information is 16 kHz, one channel, and 16 bits.

Requirements

pip install chainer
pip install hgtk
pip install python-Levenshtein
pip install typeguard
pip install librosa
pip install configargparse
pip install torch_complex
pip install pytorch_wpe
pip install humanfriendly

conda install editdistance

Total progress

1) Preprocessing the dataset

2) Models

Encoding

Transformer model

3) Hyper Parameter Optimization

We used bayesian optimization to find optimal beam size, penalty score, and CTC weight to inference model

Results

We achieve the 4.5 CER in NHN ASR hackerthon dataset (not publicly available now)

1) Our Dataset

Dataset	number of dataset	CER
Validation	-	5.2
Test	-	5.5

2) Test Samples

Dataset	number of dataset	CER
Test	-	14.5

Usage

Input data folder samples are in data folder.

#input_data_path : folder
#output_data_path : txt file
python evaluation --input_dir "input_data_path" --output_dir "output_data_path"