Incorporating Weak Statistics for Low-Resource Language Modeling
Embargo until
Date
2014-02-20
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Johns Hopkins University
Abstract
Automatic speech recognition (ASR) requires a strong language model to guide the acoustic model and favor likely utterances. While many tasks enjoy billions of language model training tokens, many domains which require ASR do not have readily available electronic corpora.The only source of useful language modeling data is expensive and time-consuming human transcription of in-domain audio. This dissertation seeks to quickly and inexpensively improve low-resource language modeling for use in automatic speech recognition.
This dissertation first considers efficient use of non-professional human labor to best improve system performance, and demonstrate that it is better to collect more data, despite higher transcription error, than to redundantly transcribe data to improve quality. In the process of developing procedures to collect such data, this work also presents an efficient rating scheme to detect poor transcribers without gold standard data.
As an alternative to this process, automatic transcripts are generated with an ASR system and explore efficiently combining these low-quality transcripts with a small amount of high quality transcripts. Standard n-gram language models are sensitive to the quality of the highest order n-gram and are unable to exploit accurate weaker statistics. Instead, a log-linear language model is introduced, which elegantly incorporates a variety of background models through MAP adaptation. This work introduces marginal class constraints which effectively capture knowledge of transcriber error and improve performance over n-gram features.
Finally, this work constrains the language modeling task to keyword search of words unseen in the training text. While overall system performance is good, these words suffer the most due to a low probability in the language model. Semi-supervised learning effectively extracts likely n-grams containing these new keywords from a large corpus of audio. By using a search metric that favors recall over precision, this method captures over 80% of the potential gain.
Description
Keywords
language modeling, automatic speech recognition, semi-supervised