5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

Creating Speaker Independent HMM Models for Restricted Database Using STRAIGHT-TEMPO Morphing

Alexandre Girardi, Kiyohiro Shikano, Satoshi Nakamura

NAIST - Nara Institute of Science and Technology, Japan

In speaker independent speech recognition, one problem we often face is the insufficient database for training. This problem is even more serious for children database. Besides, adult data when used as children data is affected by differences in pitch and spectral frequency stretch that affects recognition. In this paper, as an approach to solve the above problem, we applied STRAIGHT-TEMPO algorithm to morph adult data towards children data, in order to construct more robust HMM acoustic models, as well as to study the effect of a combined change in the pitch and spectral frequency stretch of the original utterances in the database. Using the morphed database, we analyzed the level of improvement that can be obtained, in terms of recognition rate, compared with non morphed data.

Full Paper

Bibliographic reference.  Girardi, Alexandre / Shikano, Kiyohiro / Nakamura, Satoshi (1998): "Creating speaker independent HMM models for restricted database using STRAIGHT-TEMPO morphing", In ICSLP-1998, paper 0687.