Digital Speech Processing : Automatic Speech Recoginition with LibriSpeech dataset

How to train

# Set the environment
0. Virtual environment (ex: conda) is recommended 
1. ' pip install -r requirements.txt' or install : numpy, pandas, python-Levenshtein, librosa, numba, matplotlib
2. Install pytorch considering your CUDA version : https://pytorch.org/get-started/locally/
3. ' unzip Libri_data.zip -d ./DSP_project/ ' # unzip dataset
4. Set PATHs in train.py and test.py (+ set data path at ./Libri_Data/*.csv files)

# train and test
5. ' python DSP_ASR_LibriSpeech/train.py '
6. ' python DSP_ASR_LibriSpeech/test.py '  

Recommended Environmet

• Ubuntu 20.04.3 LTS
• CUDA Version: 11.5
• Python 3.8

Dataset

: Part of Libri Speech Clean data 360 (label length is under 150) - Download it from online

Base Model

: Location based Attention - CER 36.318% WER 73.181% at 160 Epochs - CER 31.254% WER 66.985% at 200 Epochs

Performance metric

: Word Error Rate = (S + I + D) / N - S : # of substitutions - I : # of deletions - D : # of insertions - N : # of words in the target label

Reference

• specaugmentation code : https://github.com/SeanNaren/deepspeech.pytorch
• Model (CLOVA Call) : https://github.com/clovaai/ClovaCall

For Beginners

Keywords, good to study

• Listen, attend and spell
• SpecAugment
• Learning scheduler
• teacher forcing

Good reference

• Paperwithcode for LibriSpeech https://paperswithcode.com/sota/speech-recognition-on-librispeech-test-clean

For students taking DSP class

submit as zip file, including ..

• All code files you uesd for training
• One best trained model(.pth) and test reult
• Report as pdf (detail will be announced)