Search

Distributed speech recognition DSR splits the processing of data between a mobile device and a network server. In the front-end, features are extracted and compressed to transmit over a wireless channel to a back-end server, where the incoming stream is received and reconstructed for recognition tasks. In this paper, we propose a feature compression algorithm termed suppression by selecting wavelets SSW to achieve the two main goals of DSR: Minimizing memory and device requirements while also maintaining or even improving the recognition performance. The SSW approach first applies the discrete wavelet transform DWT to filter the incoming speech feature sequence into two temporal subsequences at the client terminal. Feature compression is achieved by keeping the low modulation frequency subsequence while discarding the high frequency counterpart. The low-frequency subsequence is then transmitted across the remote network for specific feature statistics normalization. Wavelets are favorable for resolving the temporal properties of the feature sequence, and the down-sampling process in DWT achieves data compression by reducing the amount of data at the terminal prior to transmission across the network. Once the compressed features have arrived at the server, the feature sequence can be enhanced by statistics normalization, reconstructed with inverse DWT, and compensated with a simple post filter to alleviate any over-smoothing effects from the compression stage. Results on a standard robustness task Aurora-4 and on a Mandarin Chinese news corpus showed SSW outperforms conventional noise-robustness techniques while also providing nearly a 50% compression rate during the transmission stage of DSR systems.

We propose a multi-style learning ( multi-style training + deep learning ) procedure that relies on deep denoising autoencoders (DAEs) to extract and organize the most discriminative information in a training database. Traditionally, multi-style training procedures require either collecting or artificially creating data samples (e.g., by noise injection or data combination) and training a deep neural network (DNN) with all of these different conditions. To expand the applicability of deep learning, the present study instead adopts a DAE to augment the original training set. First, a DAE is utilized to synthesize data that captures useful structure in the input distribution. Next, this synthetic data is combined and mixed within the original training set to exploit the powerful capabilities of DNN classifiers to learn the complex decision boundaries in heterogeneous conditions. By assigning a DAE to synthesize additional examples of representative variations, multi-style learning makes class boundaries less sensitive to corruptions by enforcing back-end DNNs to emphasize on the most discriminative patterns. Moreover, this deep learning technique mitigates the cost and time of data collection and is easy to incorporate into the internet of things (IoT). Results showed these data-mixed DNNs provided consistent performance improvements without even requiring any preprocessing on the test sets.

ESSEM allowed unsupervised model adaptation with only one testing utterance.ATG mapping function enhances ESSEM framework with sufficient adaptation statistics.Adaptation performance determined by appropriate selection of ensemble models.Over-fitting issues handled with optimization processes: MAP, MS, and CS.Imposing constraints on ATG allows flexible extension to LR, BC, LCB, LC, and BF. The ensemble speaker and speaking environment modeling (ESSEM) framework was designed to provide online optimization for enhancing workable systems under real-world conditions. In the ESSEM framework, ensemble models are built in the offline phase to characterize specific environments based on local statistics prepared from those particular conditions. In the online phase, a mapping function is computed based on the incoming testing data to perform model adaptation. Previous studies utilized linear combination (LC) and linear combination with a correction bias (LCB) as simple mapping functions that only apply one weighting coefficient on each model. In order to better utilize the ensemble models, this study presents a generalized affine transform group (ATG) mapping function for the ESSEM framework. Although ATG characterizes unknown testing conditions more precisely using a larger amount of parameters, over-fitting issues occur when the available adaptation data is especially limited. This study handles over-fitting issues with three optimization processes: maximum a posteriori (MAP) criterion, model selection (MS), and cohort selection (CS). Experimental results showed that ATG along with the three optimization processes enabled the ESSEM framework to allow unsupervised model adaptation using only one utterance to provide consistent performance improvements.

For cochlear implant users, combined electro-acoustic stimulation (EAS) significantly improves the performance. However, there are many more users who do not have any functional residual acoustic hearing at low frequencies. Because tactile sensation also operates in the same low frequencies (

Across bilateral cochlear implants, contralateral threshold shift has been investigated as a function of electrode difference between the masking and probe electrodes. For contralateral electric masking, maximum threshold elevations occurred when the position of the masker and probe electrode was approximately place-matched across ears. The amount of masking diminished with increasing masker-probe electrode separation. Place-dependent masking occurred in both sequentially implanted ears, and was not affected by the masker intensity or the time delay from the masker onset. When compared to previous contralateral masking results in normal hearing, the similarities between place-dependent central masking patterns suggest comparable mechanisms of overlapping excitation in the central auditory nervous system.

Alternative features were derived from extracted temporal envelope bank (TBANK). These simplified temporal representations were investigated in alignment procedures to generate frame-level training labels for deep neural networks (DNNs). TBANK features improved temporal alignments both for supervised training and for context dependent tree building.

This study proposes a polynomial based feature transferring (PFT) algorithm for acoustic feature conversion. The PFT process consists of estimation and conversion phases. The estimation phase aims to compute a polynomial based transfer function using only a small set of parallel source and target features. With the estimated transfer function, the conversion phase converts large sets of source features to target ones. This study evaluates the proposed PFT algorithm using a robust automatic speech recognition (ASR) task on the Aurora-2 database. The source features were MFCCs with cepstral mean and variance normalization (CMVN), and the target features were advanced front end features (AFE). Compared to CMVN, AFE provides better robust speech recognition performance but requires more complicated and expensive cost for feature extraction. By PFT, we intend to use a simple transfer function to obtain AFE-like acoustic features from the source CMVN features. Experimental results on Aurora-2 demonstrate that the PFT generated AFE-like features that can notably improve the CMVN performance and approach results achieved by AFE. Furthermore, the recognition accuracy of PFT was better than that of histogram equalization (HEQ) and polynomial based histogram equalization (PHEQ). The results confirm the effectiveness of PFT with just a few sets of parallel features.

Conventionally, the maximum likelihood (ML) criterion is applied to train a deep belief network (DBN). We present a maximum entropy (ME) learning algorithm for DBNs, designed specifically to handle limited training data. Maximizing only the entropy of parameters in the DBN allows more effective generalization capability, less bias towards data distributions, and robustness to over-fitting compared to ML learning. Results of text classification and object recognition tasks demonstrate ME-trained DBN outperforms ML-trained DBN when training data is limited.

Residual acoustic hearing can be preserved in the same ear following cochlear implantation with minimally traumatic surgical techniques and short-electrode arrays. The combined electric-acoustic stimulation significantly improves cochlear implant performance, particularly speech recognition in noise. The present study measures simultaneous masking by electric pulses on acoustic pure tones, or vice versa, to investigate electric-acoustic interactions and their underlying psychophysical mechanisms. Six subjects, with acoustic hearing preserved at low frequencies in their implanted ear, participated in the study. One subject had a fully inserted 24 mm Nucleus Freedom array and five subjects had Iowa/Nucleus hybrid implants that were only 10 mm in length. Electric masking data of the long-electrode subject showed that stimulation from the most apical electrodes produced threshold elevations over 10 dB for 500, 625, and 750 Hz probe tones, but no elevation for 125 and 250 Hz tones. On the contrary, electric stimulation did not produce any electric masking in the short-electrode subjects. In the acoustic masking experiment, 125–750 Hz pure tones were used to acoustically mask electric stimulation. The acoustic masking results showed that, independent of pure tone frequency, both long- and short-electrode subjects showed threshold elevations at apical and basal electrodes. The present results can be interpreted in terms of underlying physiological mechanisms related to either place-dependent peripheral masking or place-independent central masking.

The aim of this study is to devise a computational method to predict cochlear implant (CI) speech recognition. Here, we describe a high-throughput screening system for optimizing CI speech processing strategies using hidden Markov model (HMM)-based automatic speech recognition (ASR). Word accuracy was computed on vocoded CI speech synthesized from primarily multi-channel temporal envelope information. The ASR performance increased with the number of channels in a similar manner displayed in human recognition scores. Results showed the computational method of HMM-based ASR offers better process control for comparing signal carrier type. Training-test mismatch reduction provided a novel platform for reevaluating the relative contributions of spectral and temporal cues to human speech recognition.