1. Speech Recognition Software Modules for Multimedia Language Learning Tools

This project develops a speech recognition engine that is designed to be integrated into speech recognition and language learning systems. The engine consists of two classes: CWaveAudio and CSphRec. The first class is for collecting audio data from audio devices and the second class is for recognizing the audio data. The engine is based on the continuous density hidden Markov models, which is currently the most effective models for speech recognition. The consultancy service also provides examples of how to use the engine for isolated-word speech recognition.

A demonstration of speech recognition using the software modules
 

2. Speech Codec for Multimedia Applications

This project is to assist a company to develop a CELP-based speech decoding chip. The flow of the algorithms, its bit allocation patterns and software implementation are broken down into manageable pieces. Software-based encoders/decoders are also provided to enable the company to evaluate and compare CELP codecs and LPC-10 codecs. 

3. CELP-Based Encoders for Consumer Electronics Products

This project is to assist a company to develop a CELP-based speech encoding chip. The flow of the algorithms, its bit allocation patterns and software implementation are broken down into manageable pieces. Software-based encoders/decoders are also provided. 

4. Measurement of Xylophones’ Frequency Spectra

Objective: To measure the frequency spectra of 9 sets of xylophones, each containing eight metal plates. To determine whether the tones produced by the plates meet the required specification.

The measurement was conducted in a quiet room using a high quality close talk dynamic microphone (Stedman N90), a Sony DTC-ZA5ES DAT deck, and a PC running a signal processing package. The analog signals were produced by hitting the metal bars of the xylophone with the plastic stick provided. The signals were picked up by the microphone and sampled by the DAT deck at a rate of 48kHz with 16 bits per samples. The high dynamic range and sampling rate of the DAT deck enable frequency spectra ranging from 20Hz to 20kHz to be measured, with a frequency resolution of +/-0.98Hz. To prevent the sound generated by the mechanical contact between the metal bar and the plastic stick from interfering the measurement, the first 0.1 sec. of the signals was ignored. Each metal bar was hit 25 times to obtain 25 spectra from which the mean fundamental frequency and standard deviation were calculated. The spectra are shown in Section 3 of this document. 
 

Last update: 7 April 2003

M.W. Mak's Homepage