Audio and Voice Coder Demonstration |
||
Dr. C.F. Chan has developed many voice and audio coders, some of them are described in the following pages. |
||
| Code-Excited Sample-by-Sample Gain Adaptive Trellis Coding This is a high quality audio coder with compression ratios between 4 to 16. The decoding complexity is extremely low which is particularly suitable for applications such as digital wireless headphone in which the processing resource is severely limited.
Original (16 bit per sample), Coded at 3 bit per sample, Coded at 2 bit per sample, Coded at 1.5 bit per sample |
||
Sample-by-Sample Adaptive Differential Vector Quantization (SADVQ)SADVQ coder is a low complexity, regular structure audio coder which can operate at various bit rates (from 1 bit per sample to 3 bits per sample) depending on the sampling frequency. It can also be used to code wideband music signal. The advantage of SADVQ is its simple encoder/decoder structure which allows ASIC implementation with low gate counts (less than 5000 gates). This algorithm has been licensed to company for making digital telephone answering devices and music synthesis chips. Female (185 kB) : Original
|
||
Multi-Pulse Excited Run-Length Vector Quantization (MPE-RLVQ)MPE-RLVQ algorithm is based on the analysis-by-synthesis technique to code speech signal. A linear predictive synthesis filter is used to model the human vocal tract response. A sequence of pulses is used to excite the LP filter to generate high-quality speech. MPE-RLVQ algorithm is a relatively low complexity algorithm which has been utilized in CommonTalk, a digital simultaneous voice over data modem, employing only one TMS320C25 DSP.
|
||
Harmonic Excited Linear Prediction (HELP)HELP coder is a relatively high-complexity speech coder. It can operate at an amazingly low bit rate of 2.0 kbit/s with high speech quality. Recently, the HELP speech coder was improved to code speech at 960 bit/s and the speech quality achieved is even better than the old version of a higher rate. The HELP algorithm is based on using a very efficient representation of voice excitation model, the harmonic model. This HELP coder has been implemented in real-time (full duplex) on TI TMS320C31 digital signal processor.
|
||
Self-Orthogonal Code Excited Linear Prediction (SO-CELP)SO-CELP coder employs the code excitation technique widely used in many modern speech coders. Conventional CELP coder demands very high computational power in order to search for the excitation codeword in the stochastic codebook which typically requires high-speed DSP for real-time implementation. SO-CELP coder employs a specially designed codebook where each codevector in the codebook is self- orthongonal. As a result, codebook search can be speed up tremendously. A 4 kbit/s SO-CELP coder has been successfully implemented on a digital telephone answering device (DTAD) using Motorola DSP56166 DSP.
|
||
Odd-Symmetry Crosscorrelation Code Excited Linear Prediction (OSC-CELP)OSC-CELP is one of most advanced speech coding algorithms available today. Its employs the vector-sum excitation technique utilized in digital mobile phone systems in North America and Japan. By designing the basis codevectors in the vector sum codebook to have a property that the crosscorrelation of any pair of vectors is odd symmetry, the optimum codevector can be easily determined without the need of codebook search. Therefore, OSC-CELP can be implemented on relatively low-speed DSPs such as TMS320C25. A 4.8 kbps OSC-CELP decoder has been ported on Microchip's PIC17C42A microcontroller. A demo board was built which contains 1 MB of EPROM and can store half an hour of high-quality speech. The targeted applications of the decoder board is for low-cost voice announcement systems, electronic toys and dictionary.
|
| Female (185 kB) : 7.2 kbps, 5.6 kbps, 4.8 kbps | Real time 7.2 kbps In this demo, a 7.2 kbps bit stream is downloaded to a Java applet decoder and the compressed code is then decoded and playbacked in real-time * Please install Microsoft VM (Recommanded) or Java VM if your PC or your browser doesn't
have any kind of VM. |
LPC-to-HELP Converter
This converter is based on an innovative algorithm for converting LPC coded speech (eg. TI 50C11 D6 format) to HELP format for high-quality speech synthesis. It is well known that LPC speech sounds very mechanic and requires extensive manual editing in order to achieve reasonable good quality. By converting LPC coded speech to HELP parameters and using a harmonic model for speech synthesis, high quality speech can be generated while retaining the old speech data bases coded in LPC format. A direct replacement of LPC speech synthesis chip is possible either for speech storage applications such as those used in educational toys or for text-to- speech synthesis applications.
Demonstration (Wave File)
Female (185 kB) : LPC coded LPC synthesized, LPC coded HELP synthesized
Male (130 kB) : LPC coded LPC synthesized, LPC coded HELP synthesized
Isolated-word Speech Recognizer
This speech recognizer was designed for speaker-dependent small vocabulary task with low complexity. The technique employed is based on semi-continuous hidden Markov model (HMM). For demo please send me an email.
For more information please contact
Dr. C.F. Chan at
Tel: (852) 27887758
Fax: (852) 27844294