Speaker normalization by means of paralinguistic transformations

Paralinguistic transformations (PT) change the information about a speaker's pitch, age and sex in a speech signal and leave the linguistic contents invariant. PTs are spectrum transformations and can be used to normalize speech, i.e. to compensate for variations in the speech signal due to differently shaped speech organs across speakers. The major such variations are caused by different vocal tract lengths that are directly related to the speaker's pitches. The corresponding PTs are direct functions of the fundamental frequency F0.

The focus of the project is on the following issues:

• exploration of various F0-based PTs known from acoustic phonetics and the evaluation of their performance with a standard speech recognition system. F0-based PTs allow an immediate adaption to a new speaker which is an advantage over commonly used normalization schemes.
• development of a new feature extraction method that locates peaks shaped by the formants in the spectrum by a pattern matching approach. Spectral peaks are supposed to be the most robust information-conveying acoustical features in a noisy environment. The new feature extraction method will allow a more direct evaluation of the normalization schemes by means of PTs.

More information can be found in [Gla03].

Supported by: This project was partly supported by NCCR IM2.