Recent Progress in Robust Vocabulary-Independent Speech Recognition

Recent Progress in Robust Vocabulary-Independent Speech Recognition Hsiao-Wuen Hon and Kai.Fu Lee School of Computer Science Carnegie Mellon University Pittsburgh, Pennsylvania 15213 Abstract This paper reports recent efforts to improve the performance of CMU's robust vocabulary-independent (VI) speech recog- nition systems on the DARPA speaker-independent resource management task. The improvements are evaluated on 320 sentences that randomly selected from the DARPA June 88, February 89 and October 89 test sets. Our first improvement involves more detailed acoustic modeling. We incorporated more dynamic features computed from the LPC cepstra and reduced error by 15% over the baseline system. Our second improvement comes from a larger training database. With more training data, our third improvement comes from a more detailed subword modeling. We incorporated the word bound- ary context into our VI subword modeling and it resulted in a 30% error reduction. Finally, we used decision-tree allophone clustering to find more suitable models for the subword units not covered in the training set and further reduced error by 17%. All the techniques combined reduced the VI error rate on the resource management task from 11.1% to 5.4% (and from 15.4% to 7.4% when training and testing were under dif- ferent recording environment). This vocabulary-independent performance has exceeded our vocabulary-dependent perfor- mance. Introduction As speech recognition flourishes and new applications emerge, the demand for vocabulary-specific training will become the bottleneck in building speech recognizers. If successful, a vocabulary-independent (VI) speech recogni- tion system trained on a large database alleviates the te- dious vocabulary-specific training process. We have previ- ously demonstrated the feasibility of vocabulary-independent speech recognition systems [4, 5]. Although the vocabulary- independent results improved as the training data increases, the best vocabulary-independent result previously reported was still about 30% worse than the vocabulary-dependent (VD) result. In this paper, we will report recent efforts to fur- ther improve CMU's robust vocabulary-independent speech recognition systems on the DARPA speaker-independent re- source management task. Our first improvement involves the incorporation of more dynamic features in theacoustic front-endprocessing [7]. Our previous vocabulary-independent experiments have used only first order differenced cepstra and power. Here, we add second order differenced cepstra and power. We also incorporateboth 40 msec and 80 msec differenced cepstraa. These new features yielded a 15% error rate reduction, about the same as was achieved on vocabulary-dependent tasks [7]. Our second improvement involves the collection of more general English data, from which we can model more pho- netic variabilities, such as the word boundary context. Our experiment shows that adding 5,000 sentences to an original 15,000 sentence training set gives only a 3% error reduction. In this experiment, the set of models was fixed. Next, we incorporated word boundary context into our VI subword modeling, which resulted in a surprising 30% error reduction. This compares with only a 20% error reduction obtained in the vocabulary-dependent case. In the past, it has been argued that between-word triphones may be learning grammatical constraints instead of modeling acoustic vari- ations. This result shows the contrary, since in vocabulary- …

Content truncated to 500 words