Thumbnail Image

Using the beat histogram for speech rhythm description and language identification

Lykartsis, Athanasios; Weinzierl, Stefan

FG Audiokommunikation

In this paper we present a novel approach for the description of speech rhythm and the extraction of rhythm-related features for automatic language identification (LID). Previous methods have extracted speech rhythm through the calculation of features based on salient elements of speech such as consonants, vowels and syllables. We present how an automatic rhythm extraction method borrowed from music information retrieval, the beat histogram, can be adapted for the analysis of speech rhythm by defining the most relevant novelty functions in the speech signal and extracting features describing their periodicities. We have evaluated those features in a rhythm-based LID task for two multilingual speech corpora using support vector machines, including feature selection methods to identify the most informative descriptors. Results suggest that the method is successful in describing speech rhythm and provides LID classification accuracy comparable to or better than that of other approaches, without the need for a preceding segmentation or annotation of the speech signal. Concerning rhythm typology, the rhythm class hypothesis in its original form seems to be only partly confirmed by our results.