Introduction
In HMM-based Text-to-Speech (TTS), modeling the fundamental frequency (F0) is crucial for natural-sounding speech. A core component of F0 modeling is the voiced/unvoiced (v/u) decision. Human speech alternates between voiced segments (where vocal cords vibrate, like vowels) and unvoiced segments (like the "s" in "snake").
Standard Multi-Space Distribution HMMs (MSD-HMMs) typically make these decisions independently for each state. This "memoryless" approach often leads to rapid, unnatural flickering between voiced and unvoiced states, resulting in audible glitches and a "buzzy" voice quality. This paper introduces a "minimum v/u error" approach to improve naturalness by moving away from state-by-state decisions toward sequence-based optimization.
Proposed Method
To address the limitations of local, state-based decisions, the authors propose a framework that replaces the standard process with a sequence-based optimization. The method identifies segments where a transition is expected based on linguistic priors (e.g., known Mandarin phoneme properties) and then finds the single optimal switching point that minimizes the total error across that segment.
Instead of making a binary choice at every frame based on local probability, the algorithm calculates accumulated v/u posterior probabilities. For a given sequence (either Unvoiced-to-Voiced or Voiced-to-Unvoiced), it searches for the single transition point that minimizes the sum of errors for the entire segment.
By forcing a single transition point per segment, the system enforces physical constraints consistent with human speech production, leading to a much more stable and consistent F0 trajectory.
Experimental Setup
The proposed method was evaluated using a high-quality Mandarin corpus consisting of broadcast news style speech from a professional female radio announcer. The dataset included 1,000 sentences for training and 50 for testing. Features included 40th order Line Spectral Pairs (LSPs) and F0 extracted via the STRAIGHT system, refined with a voting method to ensure high-quality training data.
Key Results
The sequence-based optimization yielded significant improvements in both objective and subjective performance metrics.
Objective Performance
- Swapping Errors: Voiced-to-unvoiced (v → u) errors were reduced from 3.7% in the baseline to 2.0%.
- Intra-phone Stability: Erroneous switching within a single voiced phone was cut in half, from 8% to 4%.
Subjective Evaluation
In AB preference tests, where listeners compared audio clips from both systems, the new approach was preferred 72% of the time, compared to only 22% for the baseline system.
As shown in the F0 trajectories above, the baseline system often exhibits sharp drops to zero (unvoiced) in the middle of voiced segments (e.g., around frame 40). The proposed approach maintains a smooth, natural contour that closely tracks the "natural" reference speech.
Conclusion
The primary insight of this work is that the statistical independence of HMM states is a weakness for F0 generation. By treating the v/u decision as a search for a transition point across a sequence rather than isolated frames, the researchers aligned the mathematical model with the physical constraints of human speech. While some errors persist near silence boundaries, this method represents a significant step forward in reducing the "buzzy" artifacts common in HMM-based synthesis.
Summary updated