An information theoretical investigation into the distribution of phonetic information across the auditory spectrogram

✍ Scribed by Andrew Morris; Jean-Luc Schwartz; Pierre Escudier

• Publisher: Elsevier Science
• Year: 1993
• Tongue: English
• Weight: 790 KB
• Volume: 7
• Category: Article
• ISSN: 0885-2308
• DOI: 10.1006/csla.1993.1006

✦ Synopsis

In order to achieve real-time performance, the spatio-temporal resolution of preprocessed data entering typical speech recognition systems is limited to a level which is approximately three orders of magnitude less than that required to avoid significant loss of speech information. The problem of reducing high-resolution data to a manageable level while minimizing the loss of speech information is therefore a key issue in the design of improved recognition systems. When we look at the natural auditory system, one of the first processes applied to the auditory nerve signal is the detection of sudden onsets and offsets in signal energy such as those which are associated with consonantal closure and burst release. Depending on the degree to which such events can successfully locate concentrations of information for stop consonants and other phonemes, on/off detectors could provide a basis for reducing the data-rate in the output from high-resolution auditory models while preserving essential speech information. In this paper we use the information theoretic measure of mutual information to investigate the distribution of phonetic information across the on/off aligned auditory spectrogram for a corpus of vowel-plosive-vowel utterances. Automatic recognition is then used to test to what extent small high information samples are sufficient for plosive discrimination.