Abstract
The divergence of the theory and practice of vocal tract length normalization (VTLN) is addressed, with particular emphasis on the role of the Jacobian determinant. VTLN is placed in a Bayesian setting, which brings in the concept of a prior on the warping factor. The form of the prior, together with acoustic scaling and numerical conditioning are then discussed and evaluated. It is concluded that the Jacobian determinant is important in VTLN, especially for the high dimensional features used in HMM based speech synthesis, and difficulties normally associated with the Jacobian determinant can be attributed to prior and scaling.
Introduction
Vocal Tract Length Normalization (VTLN) is a cornerstone technique in speech processing, designed to mitigate speaker variability caused by physiological differences. Specifically, the varying lengths of human vocal tracts—typically ranging from 13cm in women to 18cm in men—result in shifted formant frequencies. VTLN applies a warping factor ($\alpha$) to the frequency axis to normalize these features against a generic model.
While the mathematical theory of feature transformation necessitates the inclusion of a Jacobian determinant to account for space compression or expansion, a "theory-practice gap" has persisted. In Automatic Speech Recognition (ASR), many practitioners found that including the Jacobian determinant actually degraded performance, leading to its widespread omission. This paper investigates this discrepancy, arguing that the Jacobian is essential, particularly for high-dimensional features, when integrated into a proper Bayesian framework.
A Bayesian Framework for VTLN
The authors propose moving from a Maximum Likelihood (ML) estimation to a Maximum A-Posteriori (MAP) estimator. This shift allows for the formal introduction of a prior distribution on the warping factor $\alpha$.
Key Components
- Matrix Formulation: VTLN is treated as a linear transformation in the cepstral domain. This allows the Jacobian to be calculated as the determinant of a transformation matrix $A$.
- Beta Mixture Prior: A two-component beta mixture is proposed as a prior for $\alpha$. This reflects the physical reality of bimodal warping factors (male vs. female) and constrains $\alpha$ within a physically plausible range.
- Numerical Conditioning: For high-dimensional features, the log-determinant is calculated via the sum of the logs of the matrix's eigenvalues to ensure stability.
- Acoustic Scaling: A scale factor is introduced to balance the acoustic likelihood against the Jacobian, preventing the latter from dominating the estimation process.
The Impact of Dimensionality
A critical finding is that the importance of the Jacobian determinant scales with the dimensionality of the feature set. In traditional ASR, which typically uses 12-15 features, the Jacobian's effect is relatively flat. However, in modern HMM-based Speech Synthesis (TTS), which utilizes higher-order features (e.g., 39th-order), the Jacobian becomes significantly more volatile.
As shown in the figure above, the Jacobian value for 39th-order features varies much more sharply with the warping factor than for 12th-order features. Ignoring the Jacobian in high-dimensional contexts leads to "boundary spreading," where warping factors are pushed toward extreme, unrealistic values.
Results and Visual Evidence
The study demonstrates that the perceived degradation caused by the Jacobian is actually a symptom of assuming a "flat prior" (where all warping factors are equally likely). When a proper bimodal prior is applied, the Jacobian-based model matches or exceeds the performance of the "no-Jacobian" heuristic.
In ASR tests (WSJ0 dataset), the Word Error Rate (WER) improved from a baseline of 22.16% to 19.33% with the Jacobian-normalized VTLN, compared to 19.43% for the version without the Jacobian.
Visual analysis of the warping factor distributions confirms that without proper normalization or priors, estimates for different speaker groups tend to overlap or cluster at the boundaries of the search space. The proposed Bayesian approach yields distinct, physically plausible clusters for male and female speakers.
Conclusion
This research reconciles the mathematical necessity of the Jacobian determinant with the practical requirements of speech processing. By accounting for proper priors and acoustic scaling, the Jacobian determinant becomes a beneficial component of VTLN, especially for high-dimensional synthesis tasks. Future work will focus on developing principled methods for determining the acoustic scale factor, which currently requires empirical estimation.
Summary updated