Work

Efficient setup of acoustic models for large vocabulary continuous speech recognition

Year: 2015
Type: dissertation
Abstract: Automatic speech transcription systems are developed for various languages, domains,and applications. Statistical models are used in conventional automatic speech transcription systems to produce the most likely transcription. For optimal performance, large task representative databases are required for data driven parameter estimationof statistical speech transcription models. The manual creation of high quality labeled speech data for statistical model training is by far the most time consuming processand also one of the largest cost factors to be considered. In this thesis, we investigate unsupervised parameter training and refinement methods for acoustic models to reduce time and cost factors in the development of automatic speech transcription systems. First, we describe the costly creation process of manually labeled training data to setup a transcription system from scratch. In contrast, we present a cost-efficient rapidporting of an existing transcription system for a new domain or language, when the setup of a transcription system is constrained by a budget or time limit. We summarize a framework to efficiently enrich automatically generated transcription hypotheses with posterior confidence scores for different transcription details, e.g.word, pronunciation, or state confidence scores. We evaluate confidence based weightingand filtering approaches for various unsupervised acoustic model adaptation techniques. Posterior confidence scores on state level are studied for unsupervised adaptation as well as for unsupervised training of acoustic model parameters. We examine different unsupervised training scenarios where the amount and thequality of available language resources vary. We improve the accuracy of a well tuned English transcription system for European parliament speeches – trained on one hundredhours of carefully manually transcribed speech data – by applying additional automatically labeled audio recordings. In contrast to this scenario, we then bootstrap the acoustic model of a Polish transcription system from a Spanish model and iteratively refine the Polish model by exclusively using automatically labeled training data. In addition, we address the scenario of improving a Persian pronunciation lexicon by unsupervised refinement of a statistical pronunciation model. Finally, we discuss the task of developing an acoustic model for a transcription system without any manually labeled training or evaluation data analogously to the task of breaking a cipher system.
Language: English
Open Access status: green
OpenAlex ID: W1772518162
Cites: 114
Cited by:
Related to: 10