4aSC33. Generation of wideband speech from bandlimited speech by use of an autocorrelation function.

Session: Thursday Morning, December 5

Time:


Author: Jouji Suzuki
Location: Dept. of Information and Comput. Sci., Saitama Univ., 255 Shimo-okubo, Urawa, 338 Japan
Author: Kayo Nagami
Location: Dept. of Information and Comput. Sci., Saitama Univ., 255 Shimo-okubo, Urawa, 338 Japan
Author: Hiroyuki Yashima
Location: Dept. of Information and Comput. Sci., Saitama Univ., 255 Shimo-okubo, Urawa, 338 Japan

Abstract:

A high quality of speech sound is requested in speech communication circuits. This requirement, however, is not satisfied in telephone and mobile communication circuits. This paper proposes a new method to generate a lower speech signal from bandlimited speech (300--3400 Hz). A short-time autocorrelation function (ACF) is calculated on a speech waveform where the amplitude is compressed with a square root. A period is determined from the ACF. The period of the ACF is employed as an output signal. This signal is low-pass filtered and added to the original bandlimited signal. On the other hand, the higher band signal is synthesized by a pitch excitation formant synthesizer with fixed formant frequencies. The synthesized signal is filtered with BPF (3.4--7 kHz) and added to the original. Those processes are repeated with every pitch period. Short sentences uttered by two males and two females are assessed with a preference test. Synthesized speech is compared with bandlimited natural speech where the lowest frequency is varied between 0 and 300 Hz. The hearing test reveals that the proposed method provides a speech signal that is equivalent to the bandlimited signal where the limit of low frequency is 100--200 Hz.


ASA 132nd meeting - Hawaii, December 1996