14th International Congress of Phonetic Sciences (ICPhS-14) San Francisco, CA, USA August 1-7, 1999

Speech Recognition from Temporal Patterns

Sangita Sharma (1), Hynek Hermansky (1,2)

(1) Oregon Graduate Institute of Science and Technology, Portland, Oregon, USA
(2) International Computer Science Institute, Berkeley, California, USA

This paper explores the nature of linguistic information present in the temporal structure of speech. We derive TempoRAl Patterns (TRAPs) which describe the temporal evolution of different phonemes in conversational speech. The temporal spread of the TRAPs in the region preceding as well as following the occurrence of a phoneme suggests that the information regarding the phoneme is available in rather long time durations around it. We propose an approach using TRAPs for phonetic feature extraction in speech recognition instead of the conventional spectral- based features. The resulting temporal-based automatic speech recognition (ASR) system yields recognition performance which is comparable to that of the conventional spectral-based ASR system and it is inherently robust to certain types of noise.

Full Paper

Bibliographic reference.  Sharma, Sangita / Hermansky, Hynek (1999): "Speech recognition from temporal patterns", In ICPhS-14, 1661-1664.