Abstract
It is generally known that a well-designed excitation produces high-quality signals in hidden Markov model (HMM)-based speech synthesis systems. This paper proposes a novel technique for generating excitation based on waveform interpolation (WI). For modeling WI parameters, we implemented a statistical method using principal component analysis (PCA). The parameters of the proposed excitation modeling techniques can be easily combined with the conventional speech synthesis system under the HMM framework. From a number of experiments, the proposed method has been found to generate more naturally sounding speech.
Introduction
In the field of speech synthesis, Hidden Markov Model (HMM) based systems are highly valued for their flexibility and small memory footprint. However, a persistent challenge is the "robotic" or "buzzy" quality of the generated audio compared to unit-selection systems. This artifact is primarily attributed to the vocoder—the component that converts mathematical parameters into sound. Traditionally, the "excitation" signal (representing vocal cord vibration) is modeled using a simple "Pulse or Noise" (PoN) method. This binary approach is too rigid to capture the subtle, evolving nuances of human vocal cord vibration, leading to synthetic-sounding speech.
Methodology
To address the limitations of PoN models, this research introduces a sophisticated excitation model based on two pillars: Waveform Interpolation (WI) and Principal Component Analysis (PCA).
Waveform Interpolation (WI)
Instead of a generic pulse, the system extracts "Characteristic Waveforms" (CW) from the Linear Prediction (LP) residual of real speech. A CW represents a single cycle of the excitation signal. The WI algorithm ensures smooth transitions between these waveforms in adjacent frames, preventing the jerky transitions that cause audio artifacts.
PCA-based Modeling and HMM Integration
Because raw waveforms contain too much data to model directly in an HMM, PCA is applied to a large collection of CWs. This allows any complex waveform to be represented as a linear combination of a few "basis vectors" (eigenvectors). In this study, the researchers found that just 8 eigenvectors were sufficient to represent the excitation signal's magnitude.
These 8 basis vector coefficients are treated as additional features—similar to pitch or spectral envelopes—and are trained using standard HMM procedures. During synthesis, the HMM predicts these coefficients to reconstruct the CW magnitude. While the magnitude is modeled via PCA, the system uses a default phase for voiced sounds and random phase for unvoiced sounds to maintain computational efficiency.
Experimental Evaluation
The method was evaluated using a dataset of 3,200 phonetically balanced utterances from a Korean male speaker. The HMM training utilized 43 phonemes, accounting for contextual factors such as preceding/succeeding phonemes and syllable positions.
Subjective listening tests (paired comparisons) were conducted with 14 listeners, comparing the proposed method against the standard PoN model and the Two-Band Excitation (TBE) model. On a preference scale of -2 (much worse) to +2 (much better), the proposed method outperformed the baselines significantly:
- vs. PoN: +0.88
- vs. TBE: +0.64
Conclusion
The integration of PCA-based waveform interpolation into the HMM framework effectively reduces the "buzziness" of synthetic speech. By treating the excitation signal as a continuous, statistically modeled waveform rather than a binary state, the system achieves a higher degree of naturalness. Future work will focus on modeling dynamic phase variations and exploring the generalizability of the PCA basis vectors across different speakers and emotional states.
Summary updated