Abstract
Traditional Linear Predictive Coding (LPC) vocoders often suffer from "buzzing" artifacts due to oversimplified excitation models. This paper presents a novel excitation model based on pitch-scaled analysis and Principal Component Analysis (PCA) to reconstruct the speech residual. By capturing fine-grained harmonic structures and utilizing hybrid excitation, the proposed system achieves high-quality speech synthesis, particularly for female voices, rivaling state-of-the-art methods like STRAIGHT while maintaining computational efficiency.
Introduction
The "buzzy" or robotic quality of speech produced by standard LPC vocoders stems from a binary excitation model that treats speech as either a periodic pulse train or white noise. While high-quality alternatives like STRAIGHT exist, their computational complexity can hinder integration into statistical parametric synthesis systems. This research introduces a parametric representation that accurately models the harmonic structure of the speech residual within a modern synthesis pipeline.
Methodology
The proposed vocoder employs a multi-step process for residual reconstruction:
Pitch-Scaled Analysis and Spectrum Extraction
Unlike methods requiring precise glottal closure instant (GCI) detection, this approach uses analysis frames spanning a small multiple of the pitch period. The periodic component is extracted from the residual's DFT spectrum, specifically the "even lines," and normalized to a length of 128 for consistency.
Dimensionality Reduction via PCA
To optimize the representation for synthesis, Principal Component Analysis (PCA) is applied to the periodic spectrum. The researchers found significant redundancy in the spectral data, allowing for substantial compression while maintaining perceptual quality.
Hybrid Excitation and Synthesis
Aperiodicity is modeled using a Harmonic-to-Noise Ratio (HNR) to determine the Voicing Cut-off Frequency (VCO). Synthesis combines a periodic component (via Inverse DFT and Overlap-Add) with an aperiodic component (filtered white noise). The final speech is generated by passing this hybrid excitation through an all-pole LPC filter.
Experimental Results
The system was evaluated using a database of female and male speech. Analysis of 1,000 female sentences provided the pitch period distribution used to inform the model.
Subjective Quality
Subjective listening tests with 10 participants yielded a Mean Opinion Score (MOS) of 4.1 for female voices, nearly matching the 4.2 score of the STRAIGHT vocoder. Performance on male voices was lower (MOS ~3.6), likely due to differences in pitch-cycle phase or aperiodic characteristics.
Compression Resilience
The PCA-based reduction proved highly effective. Reducing the periodic spectrum from 128 to 20 eigenvectors only marginally decreased the MOS for female voices from 4.1 to 4.0, confirming that the most perceptually relevant information is captured in a low-dimensional space.
Conclusion
The pitch-scaled analysis based vocoder effectively reduces buzzing artifacts in synthetic speech. Its success, particularly with female voices, highlights the importance of detailed harmonic modeling. Future work will focus on improving male voice quality through better phase modeling and integrating the vocoder into HMM-based text-to-speech systems.
Summary updated