SRI November 1993 CSR Spoke Evaluation

In this paper we present SRI’s results on the 1993 ARPA CSR Spoke Evaluations. This evaluation used the same HMM acoustic models as those used in SRI’s hub system: gender-dependent Genonic HMM’s. The system was made robust by modifying the front end algorithms to estimate the cepstral features (the HMM models were not modified). The robust front-end used a wide bandwidth (100-6400Hz) and estimated the cepstral coefficients using a series of algorithms that had little effect on the Sennheiser features while making the secondary microphone features look more like the Sennheiser features. The decoder used SRI’s DECIPHERTM speech recognition system [1-5] with a progressive search multipass HMM system, and used the Lincoln Lab 5K NVP trigram language model. 1. SYSTEM DESIGN HIGHLIGHTS An overview of the SRI robust system design used for spokes S5, S6 and S7 is shown below: Figure 1: SRI’s Robust CSR Block Diagram SRI NOVEMBER 1993 CSR SPOKE EVALUATION Mitchel Weintraub, Leonardo Neumeyer and Vassilios Digalakis SRI International Speech Technology and Research Laboratory Menlo Park, CA, 94025 2. RESULTS HIGHLIGHTS SRI’s results on Spoke S5 demonstrated that for unknown workstation microphones, there was an overall increase in the word-error rate of 27% over the Sennheiser microphone. For the Audio-Technica Microphone in Spoke S6, there was an 8% increase in word-error rate over the Sennheiser microphone. There was no significant difference between the performance using the Sennheiser microphone and the Audio-Technica Microphone. This is the first time that no significant increase in worderror was observed when training with a high-quality close-talking microphone and testing with a secondary microphone. SRI’s experiments demonstrated that unknown microphone algorithms can outperform known-microphone algorithms. By designing speech-recognition systems that use information about many different microphones, a recognition system can be designed so that small amounts of information about a new environment will not be sufficient to improve performance. The recognition analogy is that speaker-independent system with lots of training can outperform a speaker dependent system with only limited training. For this reason, SRI’s system for Spokes S6 and S7 used our best robust-system from Spoke S5. This has led to some confusion for the P0 condition for these spokes. In summary, for Spokes S6 and S7 we did not think that it was necessary to adapt to the new microphone conditions as our microphoneindependent system was robust to channel and noise conditions. 3. GENDER/MICROPHONE SELECTION The gender selection algorithm consisted of a two stage process. The first stage was a fast initial gender decision which used a single state HMM model for each gender (one state for male speech, one state for female speech). Each state used 256 Gaussian mixtures to represent the speech features. The features used for initial gender determination was the baseline zero-mean cepstral features C1-C12 augmented with pitch information. After the initial gender selection, progressive-search word lattices [3] were generated with speech-recognition models of the initial gender. These word-lattices were used to score the input utterance with a full-HMM system of each gender. The full-HMM models were then used to make the final gender selection based on HMM probability. If the HMM models reversed the decision of the earlier classifier, then new progressive-search word-lattices were generated for this sentence.