We propose a sub-band energy based end-of-utterance algorithm that is capable of detecting the time instant when the user has stopped speaking. The proposed algorithm finds the time instant at which many enough sub-band spectral energy trajectories fall and stay for a pre-defined fixed time below adaptive thresholds, i.e. a non-speech period is detected after the end of the utterance. With the proposed algorithm a practical speech recognition system can give timely feedback for the user, thereby making the behaviour of the speech recognition system more predictable and similar across different usage environments and, noise conditions. The proposed algorithm is shown to be more accurate and noise robust than the previously proposed approaches. Experiments with both isolated command word recognition and continuous digit recognition in various noise conditions verify the viability of the proposed approach with an average proper end-of-utterance detection rate of around 94% in both cases, representing 43% error rate reduction over the most competitive previously published method.

In this paper, we present a multiresolution-based feature extraction technique for speech recognition in adverse conditions. The proposed front-end algorithm uses mel cepstrum-based feature computation of subbands in order not to spread noise distortions over the entire feature space. Conventional full-band features are also augmented to the final feature vector which is fed to the recognition unit. Other novel features of the proposed front-end algorithm include emphasis of long-term spectral information combined with cepstral domain feature vector normalization and the use of the PCA transform, instead of DCT, to provide the final cepstral parameters. The proposed algorithm was experimentally evaluated in a connected digit recognition task under various noise conditions. The results obtained show that the new feature extraction algorithm improves word recognition accuracy by 41 % when compared to the performance of mel cepstrum front-end. A substantial increase in recognition accuracy was observed in all tested noise environments at all different SNRs. The good performance of the multiresolution front-end is not only due to the higher feature vector dimension, but the proposed algorithm clearly outperformed the mel cepstral front-end when the same number of HMM parameters were used in both systems. We also propose methods to reduce the computational complexity of the multiresolution front-end-based speech recognition system. Experimental results indicate the viability of the proposed techniques.

This paper investigates the combination of vocal tract length normalisation and speaker adaptation in connected digit recognition. In particular, we focus on performing this task under a continuously varying car noise environment. Continuous supervised speaker and environment adaptation is carried out on the test data according to the Bayesian framework. The paper also evaluates various approaches to implement vocal tract length normalisation. The best performance was obtained when the normalisation was performed during both initial speaker-independent training and testing. It was also noticed that, during testing, speaker specific normalisation produced better results than utterance specific normalisation. Our experimental results on the connected digit database show that the joint approach outperforms the system in which on-line Bayesian speaker adaptation is performed on HMM mean parameters. The performance gain was particularly high with so called outlier speakers for whom adaptation is truly needed.

Part of the problems in noise robust speech recognition can be attributed to poor acoustic modeling and use of inappropriate features. It is known that the human auditory system is superior to the best speech recognizer currently available. Hence, in this paper, we propose a new two-stream feature extractor that incorporates some of the key functions of the peripheral auditory subsystem. To enhance noise robustness, the input is divided into low-pass and high-pass channels to form so-called static and dynamic streams. These two streams are independently processed and recombined to produce a single stream, containing 13 feature vector components, with improved linguistic information. Speaker-dependent isolated-word recognition tests, using the proposed front-end, produced an average 39% and 17% error rate reductions, over all noisy environments, as compared to the standard Mel Frequency Cepstral Coefficient (MFCC) front-ends with 13 (statics only) and 26 (statics and deltas) feature vector components, respectively.