Structured GMM Based on Unsupervised Clustering for Recognizing Adult and Child Speech

Abstract

Speaker variability is a well-known problem of state-of-theart Automatic Speech Recognition (ASR) systems. In particular, handling children speech is challenging because of substantial differences in pronunciation of the speech units between adult and child speakers. To build accurate ASR systems for all types of speakers Hidden Markov Models with Gaussian Mixture Densities were intensively used in combination with model adaptation techniques. This paper compares different ways to improve the recognition of children speech and describes a novel approach relying on Class-Structured Gaussian Mixture Model (GMM). A common solution for reducing the speaker variability relies on gender and age adaptation. First, it is proposed to replace gender and age by unsupervised clustering. Speaker classes are first used for adaptation of the conventional HMM. Second, speaker classes are used for initializing structured GMM, where the components of Gaussian densities are structured with respect to the speaker classes. In a first approach mixture weights of the structured GMM are set dependent on the speaker class. In a second approach the mixture weights are replaced by explicit dependencies between Gaussian components of mixture densities (as in stranded GMMs, but here the GMMs are class-structured). The different approaches are evaluated and compared on the TIDIG-ITS task. The best improvement is achieved when structured GMM is combined with feature adaptation.

Full multicondition training for robust i-vector based speaker recognition

“…Unfortunately, the test conditions are often unknown [9]. MCT handles this issue by training the system on a dataset involving as diverse conditions as possible, such as different speaking styles [10], sampling rates [11], languages [12], or speaker ages [13], among others. This generally contributes to improving performance, even if the precise test conditions were never observed before [14].…”