Kevin Kreeger, kkreeger@cs.sunysb.edu

WWW Links to Speech Compression Pages

Speech and it's digital representation

• Vocal Cords and the Vocal Tract
• Infrequent variations require modeling updates typically at 50hz
• Band-limited to 4kHz and sampled at 8kHz (vs. 16kHz standard for speech recognition)
• This representation yields 9600 bits of data per second

Speech Compression Techniques

Waveform Codecs

• PCM -> 64 kbits/s
• DPCM -> less than 64 kbits/s
• ADPCM -> 32 kbits/s

Source Codecs

• Vocoders -> 2.4 kbits/s ( barely intelligible )

Hybrid Codecs

• MPE -> 10kbits/s
• RPE -> 13kbits/s ( with better quality than MPE )
• CELP -> 4.8-16kbits/s

Standards

• G.711 u-law and A-law PCM
• G.721, G.726, G.727 ADPCM
• G.728 Low Delay CELP
• GSM (European RPE)
• Federal Standard 1016 DoD CELP

Platform Requirements for Real Time execution

• G.728 CELP -> dedicated DSP
• Dod CELP -> Sun SPARC
• GSM -> 486/66mHz
• xxPCM -> General Purpose Processor

Examples of Available RealTime Systems

• HelloWorld  PC Vocoder
• True Speech
• CVI Inc.

Examples of Downloadable Speech Compression algorithms

• G.728 CELP
• DoD CELP
• GSM
• G.711 PCM

Conclusions

• Bit Rate vs. Processing Power vs. Reproduced Sound Quality
• Free source code for CELP requires a DSP or is low quality DoD CELP
• Other free code does not meet 9.6 kbit/s goal