Introduction
Statistical parametric speech synthesis, specifically HMM-based speech synthesis (HTS), typically relies on a source-filter model. While the vocal tract filter is often well-modeled, the excitation source is frequently oversimplified as a pulse train for voiced speech and white noise for unvoiced speech. This "pulse-or-noise" model often results in "buzzy" or robotic synthesized speech because it fails to capture the intricate harmonic structure of the real glottal source. This paper proposes a more sophisticated excitation model based on the amplitude spectrum of the residual signal to enhance the naturalness of HTS.
The Proposed Excitation Model
The core of the proposed method is the detailed modeling of the residual signal's amplitude spectrum. The process begins with inverse filtering of the input speech to obtain the residual signal.
Decomposition and Modeling
The residual signal is decomposed in the frequency domain into periodic (harmonic) and aperiodic (noise) components.
- Periodic Component: Instead of a simple pulse, the model uses the amplitude spectrum of a half-pitch period. To handle the high dimensionality of this data for HMM training, the authors employ the Linde-Buzo-Gray (LBG) algorithm to generate phoneme-specific codebooks, specifically for Mandarin "finals" (vowels).
- Aperiodic Component: The noise component is simplified into four parameters representing average energy in high-frequency bands (4000–8000 Hz), preserving the low-frequency harmonic structure.
Integration with HMM-based Synthesis
The extracted periodic and aperiodic parameters are integrated into the standard HMM training framework alongside fundamental frequency ($F_0$) and spectral envelope features. During the synthesis phase, the system generates these parameters from the trained HMMs. The residual signal is then reconstructed using a zero-phase criterion and the Pitch Synchronous Overlap Add (PSOLA) method.
Experimental Evaluation
The model was evaluated using a female Mandarin corpus of approximately 2 hours. The $F_0$ contours were manually adjusted to provide a high-quality baseline.
Key Results
ABX preference tests were conducted to compare the proposed model against established benchmarks:
- Vocoder Quality: In analysis-synthesis tests, the proposed model achieved quality comparable to the high-end STRAIGHT vocoder.
- HTS Performance: When integrated into a full HMM system, the proposed model significantly outperformed the HTS-STRAIGHT baseline. Listeners preferred the proposed model for its superior ability to capture harmonic detail, reducing the "muffled" quality often associated with HMM-generated speech.
Conclusion
By focusing on the amplitude spectrum of the residual signal, the proposed excitation model significantly improves the naturalness of HMM-based speech synthesis. While some audible discontinuities remain, the approach successfully bridges the gap between simple parametric models and high-quality vocoders without requiring complex Glottal Closure Instant (GCI) detection. Future work will explore better compression techniques for the periodic spectrum and validation across other languages.
Summary updated