SPEECH COMMUNICATIONS FROM AN INFORMATION THEORY VIEWPOINT.
Abstract
Speech is treated as a sequence of discrete code elements, the phonemes. The entropy of phonemes and spoken words is examined as a function of the length of intersymbol influence and size of spoken vocabulary. Bounds for word entropy are established in certain cases. Using the known perceptual cues of the phonemes and their characteristics, a decision sequence is constructed to uniquely determine any phoneme. Using the spectral composition of these cues as reported in the literature permits the calculation of an 'information spectral density' for each phoneme. Using the information rate spectral density and the power spectral density of speech an optimum linear filter for processing speech prior to transmission through a power limited channel with additive Gaussian noise is derived. If the noise is also white and small compared to the signal, the square of the absolute value of the filter transfer function is found to be proportional to the information rate spectral density divided by the speech power spectral density. A new approach to the design of non-linear filters (amplitude limiters) for speech is also discussed.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1965
- Accession Number
- AD0612264
Entities
People
- John David Griffiths
Organizations
- Syracuse University