Discriminative Fuzzy Clustering Maximum a Posterior Linear Regression for Speaker Adaptation

Ting-Yao Hu, Yu Tsao, Lin-Shan LeeInterspeech 2012 · DOI

Introduction

Automatic Speech Recognition (ASR) systems face significant challenges due to speaker variability. Inter-speaker differences in vocal tract length and intra-speaker variations caused by factors like health or mood can severely degrade performance. Speaker adaptation techniques aim to bridge this gap by adjusting a general speaker-independent (SI) model to a specific user using a small amount of adaptation data.

While Maximum Likelihood Linear Regression (MLLR) is a standard approach, it suffers from two primary issues: overfitting in data-scarce scenarios and a lack of discriminative power between phonetically similar sounds. This paper introduces DFCMAPLR, a method that integrates discriminative fuzzy clustering with Bayesian Maximum a Posterior (MAP) estimation to address these weaknesses.

Methodology: The DFCMAPLR Framework

The proposed approach builds on the Fuzzy Clustering MLLR (FCMLLR) framework. Unlike traditional methods that use rigid class trees, FCMLLR employs fuzzy weights to allow speech sounds to belong to multiple clusters simultaneously, providing a more nuanced transformation of the model parameters.

The FCMLLR adaptation process involves iterative estimation of fuzzy weights and affine transforms to produce a speaker-adapted (SA) model.

The DFCMAPLR method refines this process through a two-stage iterative estimation:

  1. Discriminative Fuzzy Weight Estimation (FWE): Instead of standard likelihood, a Likelihood Ratio (LR) objective function is used. This forces the fuzzy weights to maximize the distance between the target sound and its competitors, enhancing the model's "sharpness" and ability to distinguish between similar phonemes.
  2. MAP-based Affine Transform Estimation (ATE): To prevent overfitting when adaptation data is limited, Maximum a Posterior (MAP) estimation incorporates prior knowledge from the SI model. This acts as a statistical anchor, ensuring stable transformations even with only a few user utterances.

Experimental Setup

The system was evaluated using the ASTTEL200 Mandarin corpus. The SI model was trained on 190 speakers (38,000 utterances). Testing was performed on 10 speakers (5 male, 5 female). Crucially, adaptation was limited to just 10 utterances per speaker, simulating a high-constraint real-world scenario where a user provides very little data.

Results and Analysis

The researchers compared the hybrid DFCMAPLR (D+MAP) approach against several baselines. The results demonstrate that combining discriminative weights with MAP-based transforms yields the best performance across all test sets.

Comparison of Phone Error Rates (PER) across different adaptation methods and speaker sets. The D+MAP method consistently achieves the lowest error.

Key findings include:

  • Baseline (No adaptation): 24.04% Phone Error Rate (PER).
  • Standard MLLR: 22.67% PER.
  • Standard FCMLLR: 22.32% PER.
  • DFCMAPLR (D+MAP): 21.67% PER.

This represents a 9.86% relative reduction in error compared to the unadapted baseline. The results confirm that "fuzzy" weights provide a more precise transformation than rigid trees, and the MAP constraint is vital for stabilizing the numerous parameters in the affine transforms when data is scarce.

Conclusion

DFCMAPLR successfully combines the separation power of discriminative learning with the robustness of Bayesian statistics. While the current study focused on supervised adaptation, future work will explore unsupervised scenarios and methods for automatically tuning the scaling factors required for different acoustic environments and noise conditions.

Summary updated

BibTeX
@inproceedings{hu2012discriminative,
  author = {Ting-Yao Hu and Yu Tsao and Lin-Shan Lee},
  title = {{Discriminative Fuzzy Clustering Maximum a Posterior Linear Regression for Speaker Adaptation}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-174},
}
Figure