On the study of very low-resource language keyword search

Pham, Van Tung; Xu, Haihua; Hai, Van; Chong, Tze Yuang; Xiao, Xiong; Chng, Eng Siong; Li, Haizhou

doi:10.1109/apsipa.2015.7415294

Cited by 4 publications

(5 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We can see that our exemplarbased system significantly outperforms other systems in term of both WER and ATWV although we use only 3-layer neural network for the score tuning module while the DNN acoustic models in row 1, 2, 3 have the 9-layer architecture. In addition to achieving better performance than other acoustic models, the exemplar-based system provides complementary information which is suitable to our system combination [34].…”

Section: Exemplar-based Acoustic Modelmentioning

confidence: 99%

See 1 more Smart Citation

Multilingual exemplar-based acoustic model for the NIST Open KWS 2015 evaluation

Hai

Xiao

et al. 2015

2015 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA)

Self Cite

View full text Add to dashboard Cite

In this paper, we investigate the use of the proposed non-parametric exemplar-based acoustic modeling for the NIST Open Keyword Search 2015 Evaluation. Specifically, kerneldensity model is used to replace GMM in HMM/GMM (Hidden Markov Model / Gaussian Mixture Model) or DNN in HMM/DNN (Hidden Markov Model / Deep Neural Network) acoustic model to predict the emission probability of HMM states. To get further improvement, likelihood score generated by the kernel-density model is discriminatively tuned by the score tuning module realized by a neural network. Various configurations for score tuning module have been examined to show that simple neural network with 1 hidden layer is sufficient to fine tune the likelihood score generated by the kernel-density model. With this architecture, our exemplar-based model outperforms the 9-layer-DNN acoustic model significantly for both the speech recognition and keyword search tasks. In addition, our proposed exemplarbased system provides complementary information to other systems and we can further benefit from system combination.

show abstract

Section: Exemplar-based Acoustic Modelmentioning

confidence: 99%

“…WER (Word Error Rate) for speech recognition and ATWV (Actual Term-Weighted Value) [34] for keyword search performance. NIST also provided two sets of keyword lists (KW list).…”

Section: A Experimental Proceduresmentioning

confidence: 99%

Multilingual exemplar-based acoustic model for the NIST Open KWS 2015 evaluation

Hai

Xiao

et al. 2015

2015 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA)

Self Cite

View full text Add to dashboard Cite

show abstract

“…The dev10h data is used for parameter tuning of KWS, and we evaluate the KWS performance on evalpart1. The DTW-based acoustic similarities are estimated using the multilingual bottleneck features trained on various corpora [28], e.g. Cantonese, Turkish.…”

Section: Nist Openkws15 Datamentioning

confidence: 99%

“…We built a morpheme-based KWS system since morpheme was shown to be effective for detecting OOV keywords [5,27,28]. We adopted the Morfessor toolkit [29] to segment both wordbased dictionary and word transcriptions into morpheme units.…”

Section: Abstractearch Systemmentioning

confidence: 99%

Rescoring Hypothesized Detections of Out-of-Vocabulary Keywords Using Subword Samples

Pham

Xiao

et al. 2016

Interspeech 2016

Self Cite

View full text Add to dashboard Cite

Rescoring hypothesized detections, using keyword's audio samples extracted from training data, is an effective way to improve the performance of a Keyword Search (KWS) system. Unfortunately such rescoring framework cannot be applied directly to Out-of-Vocabulary (OOV) keywords since there is no sample in the training data. To address this limitation, we propose two techniques for OOV keywords in this work. The first technique generates samples for an OOV keyword by concatenating samples of its constituent subwords. The second technique splits hypothesized detections into segments, then estimates the acoustic similarities between detections and subword's samples according to the similarities between segments and these samples. The similarity scores from these two techniques are used to rescore and re-rank the list of detections returned by the automatic speech recognition (ASR) systems. The experiments show that incorporating the proposed similarity scores results in a better separation between the correct and false alarm detections than using the ASR scores alone. Furthermore, experimental results on the NIST OpenKWS15 Evaluation show that rescoring with the proposed similarity scores significantly outperforms the raw ASR scores, and other methods that do not use the similarity scores, in both Maximum Term Weighted Value (MTWV) and Mean Average Precision (MAP) metrics.

show abstract

“…From such data, we extract 22 dimensional filter-bank features plus 3 dimensional pitch features to train a DNN as a feature extractor. The detailed procedure can be found in [166].…”

Section: Setup For Acoustic Similarity Estimationmentioning

confidence: 99%

Robust spoken term detection using partial search and re-scoring hypothesized detections techniques

Pham¹

View full text Add to dashboard Cite

show abstract

On the study of very low-resource language keyword search

Cited by 4 publications

References 33 publications

Multilingual exemplar-based acoustic model for the NIST Open KWS 2015 evaluation

Multilingual exemplar-based acoustic model for the NIST Open KWS 2015 evaluation

Rescoring Hypothesized Detections of Out-of-Vocabulary Keywords Using Subword Samples

Robust spoken term detection using partial search and re-scoring hypothesized detections techniques

Contact Info

Product

Resources

About