Time encoded speech (TES) proposes the transmission of speech by segmenting the waveform between real-zero crossings. Speech comprises two types of sound - voiced and unvoiced. The segment rate is low for voiced sounds and high for unvoiced sounds. Matching of the variable segment generation rate to the constant transmission rate is achieved by inserting storage buffers. Since the variation in generation rate can be large for a particularly fricative utterance, the storage buffer has to be large, and consequently, the delay in transmission is large. This thesis presents a technique which reduces the buffer size requirements and hence the delay in transmission. The technique, hybrid-TES, achieves these reductions by identifying the high segment generation regions of the speech waveform; and by storing and transmitting special symbols to indicate the reconstruction of these regions by spectrally shaped random noise.
|Date of Award