Speech Recognition MATLAB Project

ELEC-COMP 404/504: Digital Speech and Audio Processing

Project #2: Speech Recognition Due by May 22, Sunday

You will develop a small dictionary (yes/no) isolated speech recognition system in this project. It will include three phases: feature extraction, HMM training and evaluating systems’ perfor- mance with different settings. Please refer to the speech files delivered in the compressed file ti20 yes no.tgz, which are sampled at 16 kHz and available under resources section of the course web site. Deliver your source codes and your project report to the BlackBoard web-page.

Phase 1: Feature Extraction

You are expected to extract two different feature representations for the speech signal. For each representation, you will also create different discrete observation spaces using VQ. Sug- gested feature representations are PARCOR and MFCC features, which are computed by frame processing with a frame duration of 20 msec and frame shift of 10 msec. Suggested discrete observation space sizes (VQ sizes) are 16, 32, 64 and 128.

Phase 2: HMM Training

Using the HMM functions of MATLAB train models for words yes and no. In training use only observations from the train directory of the delivered data. Try at least three different state configurations (i.e. different number of states in left-to-right configuration) for each word.

Phase 3: Performance Evaluations

Using the HMM functions of MATLAB construct a two-word speech recognition system. Perform the following evaluations and discuss your results in the project report. (i) Compute word error rate (WER) of the recognition system over all the data given in test directory for all system configurations, i.e. for each feature representation with each VQ size and for each HMM configuration. (ii) Record 20 repetitions of words yes and no at 16 kHz sampling rate as you utter them to the microphone. Compute and report WER performance for your recordings.

Project 2/ti20/test/f1/nof1s1t0.wav