Multi-level acoustic modeling for automatic speech recognition

Abstract

Context-dependent acoustic modeling is commonly used in large-vocabulary Automatic Speech Recognition (ASR) systems as a way to model coarticulatory variations that occur during speech production. Typically, the local phoneme context is used as a means to define context-dependent units. Because the number of possible context-dependent units can grow exponentially with the length of the contexts, many units will not have enough training examples to train a robust model, resulting in a data sparsity problem. For nearly two decades, this data sparsity problem has been dealt with by a clustering-based framework which systematically groups different context-dependent units into clusters such that each cluster can have enough data. Although dealing with the data sparsity issue, the clustering-based approach also makes all context-dependent units within a cluster have the same acoustic score, resulting in a quantization effect that can potentially limit the performance of the context-dependent model. In this work, a multi-level acoustic modeling framework is proposed to address both the data sparsity problem and the quantization effect. Under the multi-level framework, each context-dependent unit is associated with classifiers that target multiple levels of contextual resolution, and the outputs of the classifiers are linearly combined for scoring during recognition. By choosing the classifiers judiciously, both the data sparsity problem and the quantization effect can be dealt with. The proposed multi-level framework can also be integrated into existing large-vocabulary ASR systems, such as FST-based ASR systems, and is compatible with state-of-the-art error reduction techniques for ASR systems, such as discriminative training methods. Multiple sets of experiments have been conducted to compare the performance of the clustering-based acoustic model and the proposed multi-level model. In a phonetic recognition experiment on TIMIT, the multi-level model has about 8% relative improvement in terms of phone error rate, showing that the multi-level framework can help improve phonetic prediction accuracy. In a large-vocabulary transcription task, combining the proposed multi-level modeling framework with discriminative training can provide more than 20% relative improvement over a clustering baseline model in terms of Word Error Rate (WER), showing that the multi-level framework can be integrated into existing large-vocabulary decoding frameworks and that it combines well with discriminative training methods. In speaker adaptive transcription task, the multi-level model has about 14% relative WER improvement, showing that the proposed framework can adapt better to new speakers, and potentially to new environments than the conventional clustering-based approach.