Parcimonious time frequency quantization for phoneme and speaker classification

Some specificities of the speech signal may not well be addressed by the conventional speech processing. In this paper we focuses on a parcimonious representation of speech dynamics. We propose a novel coding strategy based on speech time-frequency quantization (TFQ) using simple Allen temporal interval algebra applied on subband voicing levels. Our compressed speech representation contains only 15 integers for a speech window up to 1 s long. We evaluate the discrimination power of these features for text independent speaker (60 hours, 62 speakers) or vowel recognition (1 hour, 6 vowels) on a referenced radio broadcast news used during evaluation campaign ESTER piloted by french intelligence agency. The 30 TFQ integers code ( feature compression factor (CF) of 26) classifies 62 speakers with an error reduction of 14% relatively to the random classifier, whereas the 390 float voicing features gives similar score. This illustrates the fact that TFQ may modelize co-articulation and speaking style. Preliminary model of independent speaker vowel identification using 15 integers TFQ features (CF of 6,4) gives an error reduction of 15,1% relatively to the random classifier, whereas the 48 float voicing level gives 31%. Further works to improve our parcimonious coding are then discussed.