Speaker-adapted confidence measures for speech recognition of video lectures

Abstract

Automatic Speech Recognition applications can benefit from a confidence measure (CM) to predict the reliability of the output. Previous works showed that a word-dependent naïve Bayes (NB) classifier outperforms the conventional word posterior probability as a CM. However, a discriminative formulation usually renders improved performance due to the available training techniques. Taking this into account, we propose a logistic regression (LR) classifier defined with simple input functions to approximate to the NB behaviour. Additionally, as a main contribution, we propose to adapt the CM to the speaker in cases in which it is possible to identify the speakers, such as online lecture repositories. The experiments have shown that speaker-adapted models outperform their non-adapted counterparts on two difficult tasks from English (videoLectures.net) and Spanish (poliMedia) educational lectures. They have also shown that the NB model is clearly superseded by the proposed LR classifier.

Speaker-Adapted Confidence Measures for ASR Using Deep Bidirectional Recurrent Neural Networks

“…CM are represented by scores usually between 0 and 1 which reflect the reliability of any recognition output. Considering CM as probabilities of correctness, CE has been largely addressed as a two-class (correct or incorrect) pattern recognition problem [1], [2], [3], [4], [5]. To this effect, a binary classifier is trained to map input features to class posterior probabilities.…”

“…To this effect, a binary classifier is trained to map input features to class posterior probabilities. Under this approach, CE has been gradually improved by exploring novel features and by designing more and more accurate classifiers [1], [2], [3], [4], [5].…”

“…At the same time, adaptation of CM has shown to be very effective in improving baseline performance [4], [7], [8], [9]. This is a key point, from our point of view, especially for tasks with limited training data, since adaptation allows us to easily obtain accurate task-specific models from generic models trained on large, non-specific data sets.…”

“…To address this, we were the first to implement speaker adaptation for CM. In [4], we evaluated speaker-dependent features into an LR model with good results. Then, in a follow-up work [7], we obtained even better preliminary results by using speaker-adapted DBLSTM.…”

Reliability Evaluation of Low-voltage Switchgear Based on Maximum Entropy Principle

“…And the priori information can be used to get the priori distribution which increases the failure data. This method has been applied in most fields, such as medical system [2], web [3], electrical engineering [4], finance [5], and speech recognition [6]. While, there is no discussion about low-voltage switchgear.…”