Improving broadcast news transcription by lightly supervised discriminative training

Abstract

In this paper, we present our experiments on lightly supervised discriminative training with large amounts of broadcast news data for which only closed caption transcriptions are available (TDT data). In particular, we use language models biased to the closedcaption transcripts to recognise the audio data, and the recognised transcripts are then used as the training transcriptions for acoustic model training. A range of experiments that use maximum likelihood (ML) training as well as discriminative training based on either maximum mutual information (MMI) or minimum phone error (MPE) are presented. In a 5xRT broadcast news transcription system that includes adaptation, it is shown that reductions in word error rate (WER) in the range of 1% absolute can be achieved. Finally, some experiments on training data selection are presented to compare different methods of "filtering" the transcripts.

The Application of Hidden Markov Models in Speech Recognition

“…Unsupervised training is similar to lightly supervised training except that a general recognition language model is used, rather than the biased language model for lightly supervised training. This reduces the accuracy of the transcriptions [23], but has been successfully applied to a range of tasks [56,110]. However, the gains from unsupervised approaches can decrease dramatically when discriminative training such as MPE is used.…”

“…These transcriptions are known to be error-full and thus not suitable for direct use when training detailed acoustic models. However, a number of lightly supervised training techniques have been developed to overcome this [23,94,128].…”

Automatic Lecture Transcription Based on Discriminative Data Selection for Lightly Supervised Acoustic Model Training

“…The conventional filtering method, however, has a drawback that it significantly reduces the amount of usable training data. Moreover, it is presumed that the unmatched or less confident segments of the data are more useful than the matched segments because the baseline system failed to recognize them and may be improved with additional training [12]. Recent work by Long et al [14] proposed methods to improve the filtering by considering the phone error rate and confidence measures.…”

“…In order to increase the training data for an acoustic model, a scheme of lightly supervised training, which does not require faithful transcripts but exploits available verbatim texts, has been explored for broadcast news [10]- [12] Manuscript received February 6, 2015. Manuscript revised March 13, 2015.…”

“…Instead of simple sequence matching [10]- [12] and heuristic measure-based selection [14], [15], in this work, we propose to train a set of dedicated classifiers to select the usable data for acoustic model training. Given an aligned sequence of the ASR hypothesis and the closed caption text (and also reference text in the training phase), a set of classifiers is trained based on a discriminative model to select between the ASR result and the closed caption text, or reject both if they are not matched.…”