Introduction
Conventional Hidden Markov Model (HMM) based speech synthesis often produces audio characterized as "hoarse" or "glitchy." This degradation typically stems from errors in the voiced/unvoiced (V/UV) decision process. In standard HMM-TTS, these decisions are made independently for each state using a Multi-Space Distribution (MSD). This local approach frequently results in unnatural "breaks"—for instance, a single unvoiced state appearing in the middle of a long voiced vowel—which disrupts the natural flow of speech.
Proposed Methodology
To address these abrupt transitions, this research introduces a refinement step using a Gaussian Mixture Model (GMM) to enforce global constraints on V/UV boundaries within a phonetic unit, such as a syllable.
The proposed system operates in two main stages:
- Boundary Modeling: During training, the system analyzes a speech corpus to determine the "unvoiced percentage" for various syllables. A GMM is then trained to model the distribution of these transition points for different phonetic units.
- Synthesis Refinement: During the synthesis phase, the HMM generates multiple candidate transition points. Instead of selecting the most likely state in isolation, the system evaluates these candidates against the pre-trained GMM. The transition point with the highest global probability is selected, and a smoothing operation is applied: all states before the boundary are set to unvoiced, and all states after are set to voiced (or vice versa).
Experimental Setup
The method was evaluated using a manually verified corpus from a female Mandarin speaker, consisting of 1,000 phonetically balanced sentences (approximately 19,712 syllables). The audio was recorded at a 16 kHz sampling rate with 16-bit quantization, totaling roughly 1.45 hours of speech.
Key Results
The GMM-based refinement significantly improved the naturalness of the synthetic speech. Subjective testing yielded a Mean Opinion Score (MOS) of 3.5, compared to 3.2 for the conventional HMM-based method. In a head-to-head preference test, listeners preferred the proposed method 82.2% of the time.
Objective analysis via spectrograms confirmed that the F0 (fundamental frequency) curves became continuous in the vowel segments, eliminating the jarring breaks common in the baseline system. This is particularly beneficial for Mandarin's third tone, which often triggers V/UV errors in standard pitch detection algorithms due to its low pitch.
Conclusion and Future Work
By prioritizing natural physical constraints over isolated local probabilities, the proposed GMM refinement effectively mitigates the "hoarseness" issue in HMM-TTS. While this approach marks a significant step forward, a gap remains between the proposed method and analysis-synthesized speech. Future research may involve modeling more complex syllable structures for other languages and improving automated syllable alignment to further enhance GMM accuracy.
Summary updated