Histogram-based spectral equalization for HMM-based speech synthesis using mel-LSP

This paper describes a statistical spectral parameter emphasis technique for HMM-based speech synthesis using mel-scaled line spectral pair (mel-LSP). Spectral parameter emphasis is effective for compensating over-smoothed spectra in HMM-based speech synthesis. However, there is no conventional technique that satisfies such requirements as automatic tuning for different speakers and realtime synthesis for mel-LSP. In the proposed method, the cumulative distribution function (CDF) is calculated from the histogram of spectral parameters that are extracted from training speech data. In the same manner, CDF of spectral parameters that are generated from HMMs is constructed. Then an emphasis rule is trained so that the CDF of generated parameters equals to that of training data. After generating a spectral parameter sequence from HMMs, the spectral parameter sequence is emphasized by using the rule. Experimental results show that our proposed method improves speech quality.