Speech Recognizer using HMM Model and Baum-Welch Algorithm

Summary

This program is a speech recognizer program which is implemented using Hidden Markov Model and Baum-Welch Algorithm. The project was created for COSE362 Machine Learning Course. The program is incomplete.

Contributor

This project was contributed by Lee Chanyoung(@ku-cylee)

How to run

This project requires Python version newer than 3.6. Locations of all terminal scripts below are project root directory.

Create Virtual Environment

Create virtual environment and activate it. For Windows,

$ pip -m venv .venv
$ .venv\Scripts\activate

For macOS and Linux OS,

$ pip3 -m venv .venv
$ source .venv/bin/activate

Install Required Libraries

Install the libraries required.

$ pip install -r requirements.txt

Config .env File

Now, create ".env" file at the project directory, and specify your file directories.

• HMM_PATH: HMM model file path
• TRANSCRIPT_PATH: Transcripts file path which contains monophones
• TRAIN_DATA_PATH: Path of directory which contains the whole file with vector values for each transcripts

HMM_PATH=<hmm-model-path>
TRANSCRIPT_PATH=<transcript-path>
TRAIN_DATA_PATH=<train-data-dir-path>

For example,

HMM_PATH=~/data/hmm.txt
TRANSCRIPT_PATH=~/data/trn_mono.txt
TRAIN_DATA_PATH=~/data/trn

Location of each training data file should be able to be found by the file path specified inside the transcripts file INSIDE the TRAIN_DATA_PATH. For example, if your TRAIN_DATA_PATH is ~/data/trn, and your transcripts file contains trn/f/ac/111111.lab as vector values' file path, its actual path should be ~/data/trn/trn/f/ac/111111.txt.

Run the program

Now, you can run the program:

$ python ./src/main.py

Limitations

This project is incomplete. It is able to accumulate values for each transcripts, but is not able to calculate change values of parameters for each iteration.