The authors previously reported speaker-dependent automatic speech recognition accuracy for isolated words using eleven surface-electromyographic (sEMG) sensors in fixed recording locations on the face and neck. The original array of sensors was chosen to ensure ample coverage of the muscle groups known to contribute to articulation during speech production. In this paper we systematically analyzed speech recognition performance from sensor subsets with the goal of reducing the number of sensors needed and finding the best combination of sensor locations to achieve word recognition rates comparable to the full set. We evaluated each of the different possible subsets by its mean word recognition rate across nine speakers using HMM modeling of MFCC and co-activation features derived from the subset of sensor signals. We show empirically that five sensors are sufficient to achieve a recognition rate to within a half a percentage point of that obtainable from the full set of sensors.

We report automatic speech recognition accuracy for individual words using eleven surface electromyographic (sEMG) recording locations on the face and neck during three speaking modes: vocalized, mouthed, and mentally rehearsed. An HMM based recognition system was trained and tested on a 65 word vocabulary produced by 9 American English speakers in all three speaking modes. Our results indicate high sEMG-based recognition accuracy for the vocalized and mouthed speaking modes (mean rates of 92.1 % and 86.7% respectively), but an inability to conduct recognition on mentally rehearsed speech due to a lack of sufficient sEMG activity. Index Terms: sEMG, subvocal speech, speech recognition

We incorporate the hierarchical mixtures of experts (HME) method of probability estimation, developed by Jordan (see Neural Computation, 1994), into an HMM-based continuous speech recognition system. The resulting system can be thought of as a continuous-density HMM system, but instead of using Gaussian mixtures, the HME system employs a large set of hierarchically organized but relatively small neural networks to perform the probability density estimation. The hierarchical structure is reminiscent of a decision tree except for two important differences: each "expert" or neural net performs a "soft" decision rather than a hard decision, and, unlike ordinary decision trees, the parameters of all the neural nets in the HME are automatically trainable using the EM algorithm. We report results on the ARPA 5,000-word and 40,000-word Wall Street Journal corpus using HME models.

In this paper, we incorporate the Hierarchical Mixtures of Experts (HME) method of probability estimation, developed by Jordan [1], into an HMM(cid:173) based continuous speech recognition system. The resulting system can be thought of as a continuous-density HMM system, but instead of using gaussian mixtures, the HME system employs a large set of hierarchically organized but relatively small neural networks to perform the probability density estimation. The hierarchical structure is reminiscent of a decision tree except for two important differences: each "expert" or neural net performs a "soft" decision rather than a hard decision, and, unlike ordinary decision trees, the parameters of all the neural nets in the HME are automatically trainable using the EM algorithm. We report results on the ARPA 5,OOO-word and 4O,OOO-word Wall Street Journal corpus using HME models.