Speaker Adaptation from Limited Training in the BBN BYBLOS Speech Recognition System

Speaker Adaptation from Limited Training in the BBN BYBLOS Speech Recognition System Francis Kubala Ming-Whel Feng, John Makhoul, Richard Schwartz BBN Systems and Technologies Corporation 10 Moulton St., Cambridge, Ma. 02238 Abstract The BBN BYBLOS continuous speech recognition system has been used to develop a method of speaker adaptation from limited training. The key step in the method is the estimation of a probabilistic spectral mapping between a prototype speaker, for whom there exists a well-trained speaker-dependent hidden Markov model (HMM), and a target speaker for whom there is only a small amount of training speech available. The mapping defines a set of transformation matrices which are used to modify the parameters of the prototype model. The resulting transformed model is then used as an approximation to a well-trained model for the target speaker. We review the techniques employed to accomplish this transformation and present experimental results conducted on the DARPA Resource Management database. 1. Introduction Soon after a speech recognition system begins operation, small amounts of new speech data become available to the system as spoken utterances are successfully transcribed to text. This data is of potentially great value to the system because it contains detailed information on the current state of the speaker and the environment. The purpose of rapid speaker adaptation is to utilize such small samples of speech to improve the recognition performance of the system. Speaker adaptation offers other benefits as well. For applications which cannot tolerate the ini- tial training expense of high performance speaker-dependent models, adaptation can trade-off peak performance for rapid training of the system. For typical experimental systems being investigated today on a 1000-word continuous speech task domain, speaker-dependent training uses 30 minutes of speech (600 sentences), while the adaptation methods described here use only 2 minutes (40 sentences). For applications in which an initial speaker-independent model fails to perform adequately due to a change in the environment or the task domain not represented in the training data, adaptation can utilize an economical initial model generated from the speaker-dependent training of a single prototype speaker. Again, looking at typical systems today, speaker-independent models train on 3 1/2 hours of speech (4200 sentences), while adaptation can use a speaker-dependent model trained from 30 minutes (600 sentences). In this paper, we describe the speaker adaptive capabilities of the BBN BYBLOS continuous speech recognition system. Our basic approach to the problem is described first in section 2. Two methods for estimating the speaker transformation are described in section 3. In section 4 we present our latest results on a standard testbed database. i00 3. Methods For Computing the Transformation In 1987 [5] we reported a new algorithm for estimating a probabilistic spectral mapping between two speakers. The transformation in this method is equivalent to expanding the HMM of the prototype, replacing each state by N states and connecting them in parallel by N transitions. The transition probabilities on each of these paths axe then p(kils, ), which are the original prototype probabilities for each …

Content truncated to 500 words