ASA 128th Meeting - Austin, Texas - 1994 Nov 28 .. Dec 02

5pSP23. Speech stress classification and feature analysis using a neural network classifier.

Robust Speech Processing Lab., Dept. of Electrical Eng., Duke Univ., Box 90291, Durham, NC 27708-0291

Variability in speech production due to task induced stress contributes significantly to loss in speech processing performance. A new neural network algorithm is formulated which classifies speech under stress using the susas stress database. Five spectral parameter representations (Mel, (Delta)-Mel, (Delta)[sup 2]-Mel, Auto-Corr-Mel, and Cross-Corr-Mel cepstral) are considered as potential stressed speech relayers. Eleven stress conditions are considered which include Angry, Clear, Fast, Lombard, Loud, Normal, Question, Slow, and Soft speech. Stressed speech classification using these features is evaluated with respect to (i) pairwise class separability, (ii) an objective measure of pairwise and global feature separability, and (iii) analysis of articulatory vocal tract area functions. Perturbations in speech production under stress are reflected to varying degrees in the five speech feature representations. A neural network classifier over phoneme partitions can achieve good speaker dependent classification performance (80.6%) when stress conditions are combined into six groups of related stress domains. The implication of reliable stress classification will be discussed for applications such as monitoring speaker state, improving naturalness of speech coding, and increasing robustness of speech recognizers in adverse conditions.