Automatisk Identifiering av Inandningspauser i Spontant Tal - ett HMM/ANN-hybridsystem i Matlab

Sammanfattning

This thesis presents a system which has been implemented to satisfy a need in the

research on how speech planning interacts with syntactic and prosodic structure in

spontaneous speech. The long-term purpose of the research is to provide models for

automatic parsing of spontaneous speech and for psycholinguistical modelling of speech

production. Identification of inhalation pauses is an important step in the development

of automatic methods for spontaneous speech parsing.

Identification of inhalation pauses is considered to be a keyword-spotting speech

recognition problem. Hybrid HMM(Hidden Markov Models)/ANN(Artificial Neural

Networks) approach is applied to this problem. Method gets 90,8% in Recall, 66,4% in

Precision and 76,7% in F-score. Use of a threshold value for duration increases the Fscore

to 82,5%, therefore duration is considered to be relevant in performance

optimization. Other proposed optimization parameters are better acoustic modelling,

identification of the units causing false identifications prior to inhalation pauses

identification and production of a more appropriate spontaneous speech corpus.