Language model parameter estimation using user transcriptions

Abstract

In limited data domains, many effective language modeling techniques construct models with parameters to be estimated on an in-domain development set. However, in some domains, no such data exist beyond the unlabeled test corpus. In this work, we explore the iterative use of the recognition hypotheses for unsupervised parameter estimation. We also evaluate the effectiveness of supervised adaptation using varying amounts of user-provided transcripts of utterances selected via multiple strategies. While unsupervised adaptation obtains 80% of the potential error reductions, it is outperformed by using only 300 words of user transcription. By transcribing the lowest confidence utterances first, we further obtain an effective word error rate reduction of 0.6%.

Online Incremental Learning for Speaker-Adaptive Language Models

“…Language model adaption has been widely studied in literature. Hsu et al [1] explored the iterative use of the ASR hypotheses for unsupervised parameter estimation for n-gram language models. Similarly, [2,3,4] proposed unsupervised adaptation methods for presentation lecture speech recognition.…”

TEDxSK and JumpSK: A New Slovak Speech Recognition Dedicated Corpus

“…Moreover, the problem of frequently appearing errors in automatic transcription of lecture speech was eliminated in [12] by correction of colloquial expressions, deletion of fillers and insertion of periods using statistical post-processing techniques. Authors in [13] and [14] explore output recognition hypotheses and effectiveness of supervised and unsupervised adaptation with varying amounts of user-provided transcripts to tune the language model parameters on a lecture transcription task in English.…”

Podcastle: collaborative training of language models on the basis of wisdom of crowds

“…This paper focuses on language modeling to improve recognition performance because language modeling is another important, challenging issue in transcribing podcasts and other similar content such as YouTube video clips featuring spoken documents. In the literature, various works have been done in terms of language model (LM) adaptation for several LVCSR tasks like broadcast news [6][7], meetings [8][9], and lectures [10] [11]. These works basically took an approach that a main (or background) LM trained from large amounts of task-specific text data is adapted using an additional resource such as indomain (on-topic) text data [7][8], web-based text data [6] [11], or user-provided text data [9] [10].…”

“…In the literature, various works have been done in terms of language model (LM) adaptation for several LVCSR tasks like broadcast news [6][7], meetings [8][9], and lectures [10] [11]. These works basically took an approach that a main (or background) LM trained from large amounts of task-specific text data is adapted using an additional resource such as indomain (on-topic) text data [7][8], web-based text data [6] [11], or user-provided text data [9] [10]. In our target having the diversity of topics, however, such a large-scale task-specific corpus cannot be prepared in advance to train the background LM.…”