RESUMO
Vocoder-based speech synthesis has become a promising technique to accommodate the demands of high-quality speech analysis, manipulation, and synthesis. However, most existing works focus on how to synthesize normal human voice with high signal-to-noise ratio, neglecting individuals' pathological voice disorder in speech interaction. In this work, we propose a non-linear voice repair vocoder for pathological vowels and sentences, which takes the pathological speech as input and generates high-quality repaired speech. Our approach is specifically designed to enhance the speech quality and intelligibility for individuals with voice disorders. We employ amplitude modulated-frequency modulated (AM-FM) and Teager energy operation techniques to enhance the quality of pitch and spectral envelope. To tackle the instability and fracture problem of pitch, we present spectral tracking algorithm, which not only avoids dramatic change in the edge of voice, but also reduces the errors of half-pitch. Furthermore, we design a spectral reconstruction algorithm, which can effectively rebuild the spectral structure by energy operation to accomplish spectral envelope repair. The proposed PVR-Vocoder shows exceptional performance in pathological voice intelligibility enhancement according to various quality measures including objective indicators, subjective evaluation, and spectrum observations.
RESUMO
OBJECTIVE: Speech enhancement has become a promising technique to accommodate demands of the improvement in quality of a degraded speech signal. The main works now focus on separating normal speech from noise, but have neglected the low quality of impaired speech influenced by anomalous glottis flow. In order to effectively enhance the pathological speech, it is essential to design a separation mechanism for extracting high-dimensional timbre features and speech features separately to suppress low-dimensional noises. METHODS: In this paper, we propose an enhancement model GBNF-VAE to extract timbre efficiently by reducing anomalous airflow noise interference, and by combining the semantic features with timbre features to synthesize the enhanced speech. In particular, the bottleneck feature can characterize the timbre by the controlled number of nodes through the Golden Section method, which effectively improves computational efficiency. In addition, variational autoencoder is adopted to extract semantic features which are combined with the previous timbre features to synthesize the enhanced speech. RESULTS: Finally, spectrum observation, objective indicators and subjective evaluation all show the outstanding performance of GBNF-VAE in pathological speech quality enhancement.
