The Research About The Acoustic Model

In the information age, we have to face a large number of audio and video, and a problem that how to class the similar information and find the useful part. This is also the trend of continuous speech recognition.Broadcasting Speech contains the following features: complex background environment, speaker independent and massive amount of data. We need the system using a little data to build a base line, and then selecting some unlabeled but most informative samples to annotate them, and adding the newly transcribed samples to the training set to update the acoustic model. In this way, we can greatly reduce the number of samples transcribed. In this paper, we analyze the features of broadcasting speech, select some rules for building the broadcasting speech data base and the transcribe system. At the same time, we design an active learning algorithm and build an active learning system, then comparing the random selection and K-L distance for the initial sample selection, as well as balancing random selection other training samples, the maximum likelihood (MLE) and the posterior probability. We find out using K-L distance and the posterior probability based on confusion network select the sample can greatly reduce the sample transcribed and improve system efficiency. In addition, this article also has a comparing about vowel sound element model and phoneme element model for continuous speech recognition performance. The results shows that vowels is more suitable for Chinese acoustic modeling.