Confidence scores for acoustic model adaptation

Abstract

This paper focuses on confidence scores for use in acoustic model adaptation. Frame-based confidence estimates are used in linear transform (CMLLR and MLLR) and MAP adaptation. We show that adaptation approaches with a limited number of free parameters such as linear transform-based approaches are robust in the face of frame labeling errors whereas adaptation approaches with a large number of free parameters such as MAP are sensitive to the quality of the supervision and hence benefit most from use of confidences. Different approaches for using confidence information in adaptation are investigated. This analysis shows that a thresholding approach is effective in that it improves the frame labeling accuracy with little detrimental effect on frame recall. Experimental results show an absolute WER reduction of 2.1% over a CMLLR adapted system on a video transcription task.

Confidence- and margin-based MMI/MPE discriminative training for off-line handwriting recognition

“…The M-MPE-conf criterion can be defined in a similar manner. Note that due to the quality of the confidence metric, thresholding the confidence scores after feature selection can often result in an improved accuracy, as reported in [12]. On the one hand, the experimental results for word-confidences in Figure 12 and state-based confidences in Figure 16 suggest that the confidences are helpful, but on the other hand that the threshold itself has little impact due the proposed M-MMI-conf / M-MPE-conf approaches, which are inherently robust against outliers.…”

“…Although the prior transcript in that case contains errors, adapting on that transcript disregarding that fact generally still results in accuracy improvements [12].…”

“…Many publications in automatic speech recognition (ASR) have shown that the application of confidence scores for adaptation can improve recognition results. However, only small improvements are reported for confidence-based maximum-likelihood linear regression (MLLR) adaptation [12,31,33] or constrained-MLLR adaptation [1]. In this work, we present a novel unsupervised confidence-based discriminative model adaptation approach.…”

Towards Unsupervised Learning for Handwriting Recognition

“…In another publication Frinken [3] utilizes the concept of co-training, where two systems try to improve each other. Still all those methods rely on some small amount of annotated data [4] or a bootstrap system [5].…”

Confidence-Based Discriminative Training for Model Adaptation in Offline Arabic Handwriting Recognition

“…Many publications have shown that the application of confidence scores for adaptation can improve recognition results. However, only small improvements are reported for confidence based CMLLR adaptation [1] or MLLR adaptation [6,14,16]. In this work, we present a novel unsupervised confidence-based discriminative model adaptation approach using a modified MMI training criterion.…”