Adaptive ML-weighting in multi-band recombination of Gaussian mixture ASR

Hagen, Astrid; Bourlard, Hervé; Morris, Andrew

doi:10.1109/icassp.2001.940816

Cited by 7 publications

(3 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These ANNs give phone posterior estimates for each band, which have to be recombined (which we will perform at the frame level). Numerous methods exist for this step, ranging from simple fixed linear and non-linear combinations (via ANNs) to sophisticated methods that try to dynamically assess the reliability of the bands [1,3,9,13,14,15,16]. As this paper was partially inspired by the recent renaissance in ANN-based recognition (especially with DNNs [11]), we opted to use ANNs as recombination units as well.…”

Section: Processing and Recombination Of The Bandsmentioning

confidence: 99%

Robust Multi-Band ASR Using Deep Neural Nets and Spectro-temporal Features

Kovács

Tóth

Grósz

2014

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Abstract. Spectro-temporal feature extraction and multi-band processing were both designed to make the speech recognizers more robust. Although they have been used for a long time now, very few attempts have been made to combine them. This is why here we integrate two spectrotemporal feature extraction methods into a multi-band framework. We assess the performance of our spectro-temporal feature sets both individually (as a baseline) and in combination with multi-band processing in phone recognition tasks on clean and noise contaminated versions of the TIMIT dataset. Our results show that multi-band processing clearly outperforms the baseline feature recombination method in every case tested. This improved performance can also be further enhanced by using the recently introduced technology of deep neural nets (DNNs).

show abstract

Section: Processing and Recombination Of The Bandsmentioning

confidence: 99%

Robust Multi-Band ASR Using Deep Neural Nets and Spectro-temporal Features

Kovács

Tóth

Grósz

2014

Lecture Notes in Computer Science

View full text Add to dashboard Cite

show abstract

“…Let us now assume that the recombination is carried out based on the posterior estimates of ANNs (or the estimates of any other method we chose to process subbands with). Several methods still can be chosen for recombination, ranging from simple fixed linear combinations [29,82] to sophisticated methods that try to dynamically assess the reliability of the bands [19,69,148,167,173], including ANNs [159,160]. In the experiments we used neural nets for recombination as well, as our investigation of multiband processing was partially inspired by the recent renaissance in ANN-based recognition (especially with DNNs [84]).…”

Section: Processing and Recombination Of The Bandsmentioning

confidence: 99%

“…And once the subbands have been selected, there are several decisions to be made on processing them and on combining the results. Subbands may for example be processed using GMMs [69], HMMs [29,173], and HMM/ANN hybrids [82]. But typically this is performed via ANNs [20,71,159,160] (here we also followed this practice of using ANNs).…”

Section: Processing and Recombination Of The Bandsmentioning

confidence: 99%

Noise Robust Automatic Speech Recognition Based on Spectro-Temporal Techniques

Kovács

View full text Add to dashboard Cite

Sub-banded reconstructed phase spaces for speech recognition

Indrebo

Povinelli

Johnson

2006

Speech Communication

View full text Add to dashboard Cite

ii Preface A novel method for classification of speech phonemes, based on the combination of dynamical systems theory and filter banks, is introduced. The benefit of this approach is seen in its ability to model nonlinear characteristics of speech, something that traditional methods cannot do. The modeling tool that provides this capability is the reconstructed phase space. This space carries all the dynamical information present in the signal's underlying system. The reconstructed phase spaces used for modeling and classification of the phonemes are built using frequency sub-banded signals that are generated using a set of band-pass filters. This approach is motivated by empirical evidence that suggests humans process and recognize speech in sub-bands. Modeling and classification is performed on the sub-banded reconstructed phase spaces using Gaussian Mixture Models, and the results of the classifications for each sub-band are combined to form an overall classification. Several methods for the combination of the sub-band classifications are examined, and it is found that an un-weighted linear combination produces classification accuracies that are significantly higher than those of a classification system using reconstructed phase spaces of unfiltered signals. Results also demonstrate that the proposed phoneme classification system is competitive with state-of-the-art approaches.iii

show abstract

Adaptive ML-weighting in multi-band recombination of Gaussian mixture ASR

Cited by 7 publications

References 8 publications

Robust Multi-Band ASR Using Deep Neural Nets and Spectro-temporal Features

Robust Multi-Band ASR Using Deep Neural Nets and Spectro-temporal Features

Noise Robust Automatic Speech Recognition Based on Spectro-Temporal Techniques

Sub-banded reconstructed phase spaces for speech recognition

Contact Info

Product

Resources

About