Abstract
This paper presents a novel smoothing method for Minimum Phone Error Linear Regression (MPELR) designed to enhance speaker adaptation robustness in data-constrained environments. By integrating a prior mean distribution based on Maximum A Posteriori (MAP) estimates rather than standard Maximum Likelihood (ML) estimates, the proposed technique stabilizes the objective function. Experimental results on a large vocabulary Chinese speech recognition task demonstrate an 8.4% relative reduction in Word Error Rate (WER) for limited adaptation data while maintaining asymptotic performance parity with conventional MPELR.
Introduction
Automatic Speech Recognition (ASR) systems must account for significant speaker variability to maintain high accuracy. Speaker adaptation techniques, such as Maximum Likelihood Linear Regression (MLLR), are typically employed to transform generic acoustic models to specific users. While discriminative methods like Minimum Phone Error Linear Regression (MPELR) offer superior performance by directly minimizing phone errors, they are highly susceptible to overfitting when adaptation data is scarce.
Existing smoothing techniques for MPELR attempt to mitigate this by incorporating Maximum Likelihood (ML) statistics. However, when data is limited, these ML statistics are themselves unreliable, leading to poor smoothing performance. This paper addresses this "garbage in, garbage out" cycle by introducing a MAP-based prior.
Proposed MAP-Smoothed MPELR
The core innovation lies in the definition of the prior mean distribution used in the MPELR objective function.
- MAP Advantage: Unlike ML estimates, which rely solely on the small set of new adaptation data, MAP estimates incorporate prior knowledge from the speaker-independent (SI) model. This provides a more stable "anchor" for the smoothing process.
- Mathematical Formulation: The authors derived modified statistics for the MPELR auxiliary function. By incorporating a prior weight factor ($\tau^{MAP}$) and the speaker-independent mean ($\mu^{orig}_m$), the smoothing process effectively "pulls" the model toward a sensible default when the new data is insufficient to provide a reliable discriminative signal.
Experimental Evaluation
The method was evaluated using the National 863 High-Tech Project Chinese speech corpus. The training set comprised approximately 65 hours of speech from 140 speakers, while the test set included 6 speakers providing 260 sentences each. Acoustic models were state-tied triphone HMMs with 5,955 states and 8-component GMMs.
Performance with Limited Data
The proposed method excels in low-data regimes. With only 4 sentences (approx. 6.8 seconds) of supervised adaptation data, the WER for an MPE-trained model dropped from 8.3% using standard ML-smoothing to 7.6% with the proposed MAP-smoothing.
Asymptotic Stability
A key requirement for any adaptation technique is that it should not degrade performance as more data becomes available. The experiments confirmed that the MAP-smoothed MPELR converges to the same high performance as conventional MPELR when the adaptation set grows to 260 sentences.
Key Insights
- Robustness: The reliability of the smoothing statistics is as critical as the smoothing mechanism itself. MAP provides the necessary stability for short utterances.
- Unsupervised Adaptation: The MAP-smoothed version consistently outperformed the ML version in unsupervised modes across various data volumes, suggesting broader applicability in real-world scenarios where transcriptions are unavailable.
- Model Dependency: The benefits are most pronounced when the initial model is already trained using MPE criteria, highlighting the synergy between discriminative training and discriminative adaptation.
Conclusion
The proposed MAP-smoothed MPELR provides a robust solution for speaker adaptation when data is limited. By replacing unreliable ML priors with stable MAP estimates, the method achieves significant WER reductions in the "small data" regime without sacrificing performance in the limit. Future work may explore non-Normal prior distributions to further refine adaptation for extremely short utterances.
Summary updated