Digital Signal Processing: HMM Training and Testing with the HTK toolkit

• Construct a digital monophone recognizer with the HTK toolkit.
• See "problem_description.pdf" for more details.

Implementation

• Modified from: Baseline model and script
• In this work the following files are modified:
  • lib/proto: increase state number
  • lib/mix2_10.hed: increase number of Gaussian mixtures
  • 03_training.sh: increase training iteration
  • run.sh: an additional script that runs the whole training and testing pipeline directly
• Implementation details are specified in "report.pdf".
• An testing accuracy of 97.76% is achieved with this implementation.
• 

File Description

.
├── htk341_debian_x86_64/ ...   Compiled Binary Files of the HTK toolkit
├── src/ ...                    scripts to run the HTK toolkit, and model implementations
├── image/ ...                  Results
├── problem_description.pdf     Work spec
├── report.pdf                  Report describing this work
└── Readme.md                   This file

Environment Setup

• Download the HTK tool: < Hidden Markov Model Toolkit (HTK 3.4.1) >

$ wget http://speech.ee.ntu.edu.tw/homework/DSP_HW2-1/htk341_debian_x86_64.tar.gz
$ tar zxvf htk341_debian_x86_64.tar.gz
Place the unzip .tar folder under the same directory as ./src

• Or clone this repository:

$ git clone git@github.com:andi611/DSP_HTK_Toolkit.git

• Set path for the HTK tool with script:

└── src/
    └── bash set_htk_path.sh

• Refer to "problem_descriptiom.pdf pg.28" for more detail.

Usage

Run the whole training and testing pipeline automatically:

└── src/
    └── bash run.sh

Run each script separately:

└── src/
    ├── bash 00_clean_all.sh
    ├── bash 01_run_HCopy.sh
    ├── bash 02_run_HCompV.sh
    ├── bash 03_training.sh
    └── bash 04_testing.sh