检索结果-内蒙古大学图书馆

International Conference on Information, Intelligence, Systems and Applications (IISA)

作者： Aggelos Gkiokas Vasilis Katsouros George Carayannis Institute for Language and Speech Processing Athena — Research and Innovation Center in Information Communication and Knowledge Technologies Athens Greece School of Electrical and Computer Engineering National Technical University of Athens Athens Greece

In this paper a method for computing an audio based similarity between music excerpts is presented. The method consists of three main parts, with the first step being feature extraction, which involves the calculation of three feature sets that correspond to music timbre, rhythm and harmony. Next, for each feature set a Deep Belief Network was trained without supervision on a large music collection. The respective distances of the output units of the Deep Belief Networks between two music excerpts are computed, normalized and finally combined to form the distance measure. The proposed method was evaluated on the MIREX 2013 Audio Music Similarity task. Results are encouraging, however, they indicate that the harmonic similarity component degrades the performance.

关键词： Vectors Feature extraction Harmonic analysis Power harmonic filters Timbre Rhythm

来源：评论

学校读者我要写书评

暂无评论

DIRECT ADAPTATION OF HYBRID DNN/HMM MODEL FOR FAST SPEAKER ADAPTATION IN LVCSR BASED ON SPEAKER CODE

DIRECT ADAPTATION OF HYBRID DNN/HMM MODEL FOR FAST SPEAKER A...

引用

IEEE International Conference on Acoustics, speech and Signal processing

作者： Shaofei Xue Ossama Abdel-Hamid Hui Jiang Lirong Dai National Engineering Laboratory of Speech and Language Information Processing University of Science and Technology of China Hefei P. R. China Department of Electrical Engineering and Computer Science York University Toronto Canada

ISBN: (纸本)9781479928941

Recently an effective fast speaker adaptation method using discriminative speaker code (SC) has been proposed for the hybrid DNN-HMM models in speech recognition [1]. This adaptation method depends on a joint learning of a large generic adaptation neural network for all speakers as well as multiple small speaker codes using the standard back-propagation algorithm. In this paper, we propose an alternative direct adaptation in model space, where speaker codes are directly connected to the original DNN models through a set of new connection weights, which can be estimated very efficiently from all or part of training data. As a result, the proposed method is more suitable for large scale speech recognition tasks since it eliminates the time-consuming training process to estimate another adaptation neural networks. In this work, we have evaluated the proposed direct SC-based adaptation method in the large scale 320-hr Switchboard task. Experimental results have shown that the proposed SC-based rapid adaptation method is very effective not only for small recognition tasks but also for very large scale tasks. For example, it has shown that the proposed method leads to up to 8% relative reduction in word error rate in Switchboard by using only a very small number of adaptation utterances per speaker (from 10 to a few dozens). Moreover, the extra training time required for adaptation is also significantly reduced from the method in [1].

关键词： Loudspeakers speech recognition Error analysis HYBRID

来源：评论

学校读者我要写书评

暂无评论

SEQUENCE TRAINING OF MULTIPLE DEEP NEURAL NETWORKS FOR BETTER PERFORMANCE AND FASTER TRAINING SPEED

SEQUENCE TRAINING OF MULTIPLE DEEP NEURAL NETWORKS FOR BETTE...

引用

IEEE International Conference on Acoustics, speech and Signal processing

作者： Pan Zhou Lirong Dai Hui Jiang National Engineering Laboratory of Speech and Language Information Processing University of Science and Technology of China Hefei P. R. China Department of Electrical Engineering and Computer Science York University Toronto Canada

ISBN: (纸本)9781479928941

Recently, sequence level discriminative training methods have been proposed to fine-tune deep neural networks (DNN) after the frame-level cross entropy (CE) training to further improve recognition performance of DNNs. In our previous work, we have proposed a new cluster-based multiple DNNs structure and its parallel training algorithm based on the frame-level cross entropy criterion, which can significantly expedite CE training with multiple GPUs. In this paper, we extend to full sequence training for the multiple DNNs structure for better performance and meanwhile we also consider a partial parallel implementation of sequence training using multiple GPUs for faster training speed. In this work, it is shown that sequence training can be easily extended to multiple DNNs by slightly modifying error signals in output layer. Many implementation steps in sequence training of multiple DNNs can still be parallelized across multiple GPUs for better efficiency. Experiments on the Switchboard task have shown that both frame-level CE training and sequence training of multiple DNNs can lead to massive training speedup with little degradation in recognition performance. Comparing with the state-of-the-art DNN, 4-cluster multiple DNNs model with similar size can achieve more than 7 times faster in CE training and about 1.5 times faster in sequence training when using 4 GPUs.

关键词： Functional training Training Fasting speed cerium Neural network Cross Entropy

来源：评论

学校读者我要写书评

暂无评论

IMPROVING DEEP NEURAL NETWORKS FOR LVCSR USING DROPOUT AND SHRINKING STRUCTURE

IMPROVING DEEP NEURAL NETWORKS FOR LVCSR USING DROPOUT AND S...

引用

IEEE International Conference on Acoustics, speech and Signal processing

作者： Shiliang Zhang Yebo Bao Pan Zhou Hui Jiang Lirong Dai National Engineering Laboratory for Speech and Language Information Processing University of Science and Technology of China Hefei Anhui P. R. China Department of Electrical Engineering and Computer Science York University Toronto Canada

ISBN: (纸本)9781479928941

Recently, the hybrid deep neural networks and hidden Markov models (DNN/HMMs) have achieved dramatic gains over the conventional GMM/HMMs method on various large vocabulary continuous speech recognition (LVCSR) tasks. In this paper, we propose two new methods to further improve the hybrid DNN/HMMs model: i) use dropout as pre-conditioner (DAP) to initialize DNN prior to back-propagation (BP) for better recognition accuracy;ii) employ a shrinking DNN structure (sDNN) with hidden layers decreasing in size from bottom to top for the purpose of reducing model size and expediting computation time. The proposed DAP method is evaluated in a 70-hour Mandarin transcription (PSC) task and the 309-hour Switchboard (SWB) task. Compared with the traditional greedy layer-wise pre-trained DNN, it can achieve about 10% and 6.8% relative recognition error reduction for PSC and SWB tasks respectively. In addition, we also evaluate sDNN as well as its combination with DAP on the SWB task. Experimental results show that these methods can reduce model size to 45% of original size and accelerate training and test time by 55%, without losing recognition accuracy.

关键词： Hidden Markov models Training Neural networks speech recognition speech Computational modeling Switches

来源：评论

学校读者我要写书评

暂无评论

Speaker adaptive bottleneck features extraction for LVCSR based on discriminative learning of speaker codes

Speaker adaptive bottleneck features extraction for LVCSR ba...

引用

International Symposium on Chinese Spoken language processing

作者： Changqing Kong Shaofei Xue Jianqing Gao Wu Guo Lirong Dai Hui Jiang National Engineering Laboratory of Speech and Language Information Processing University of Science and Technology of China Iflytek Research Anhui USTC iFlytek Co. Ltd. Department of Electrical Engineering and Computer Science York University

Recently, several fast speaker adaptation methods based on the so-called speaker codes (SC) have been proposed for the hybrid DNN-HMM speech recognition model [1, 2, 3]. In these methods the target speaker features are modified to match the given speaker-independent models or the speaker-independent models are transformed towards one particular speaker based on the discriminative learning of speaker codes. Previous researches have shown that these proposed SC-based adaptation methods are very effective to adapt large DNN models using only a small amount of adaptation data. In this work, we have explored the combination of direct speaker adaptation technique in model space based on speaker codes (mSA-SC) and bottleneck features where mSA-SC is used as an extraction instrument of speaker adaptive bottleneck features. We have evaluated the proposed speaker adaptive bottleneck features extraction method in two speech recognition tasks, namely PSC Mandarin task and large scale 320-hr Switchboard task. Experimental results have verified that it is quite suitable for very large scale tasks. For example, the Switchboard results have shown that it can achieve relative 9% reduction in word error rate on an unsupervised speaker adaptation scheme.

关键词： Hidden Markov models Neural networks Adaptation models Feature extraction speech recognition Switches Training

来源：评论

学校读者我要写书评

暂无评论

Modeling predictive perceptual representation of Thai initial consonants

Modeling predictive perceptual representation of Thai initia...

引用

Oriental COCOSDA International Conference on speech Database and Assessments

作者： P. Phienphanich C. Onsuwan C. Tantibundhit N. Saimai T. Saimai Department of Electrical and Computer Engineering Faculty of Engineering Department of Linguistics Faculty of Liberal Arts Center of Excellence in Intelligent Informatics Speech and Language Technology and Service Innovation (CILS) Thammasat University Pathumthani Thailand

This work is an extension of our previous attempt to construct a spatial representation of 21 initial consonants in Thai by partitioning them into homogeneous clusters based on empirical measures of confusability and distance among phonemes. The measures were taken from perceptual identification performance of 28 listeners (seven full subjects) when stimuli were presented in noise. In present study, two methods of clustering, namely Multidimensional scaling analysis and k-means clustering were employed, yielding six different classifications and four perceptually relevant categories: intra-cluster short distance, intra-cluster long distance, inter-cluster short distance, and inter-cluster long distance. Another set of perceptual experiment (eight listeners; two full subjects) was carried out to verify the predictions. The findings reveal that the derived perceptual clusters and defined categories fit relatively well with the listeners' performance. Distinctive feature systems in phonological theory appear to provide some basis for the clustering of phonemes.

关键词： System-on-chip Noise measurement Noise Psychology speech

来源：评论

学校读者我要写书评

暂无评论

TASK-DRIVEN ATTENTIONAL MECHANISMS FOR AUDITORY SCENE RECOGNITION

TASK-DRIVEN ATTENTIONAL MECHANISMS FOR AUDITORY SCENE RECOGN...

引用

IEEE International Conference on Acoustics, speech, and Signal processing

作者： Kailash Patil Mounya Elhilali Center for Language and Speech Processing Department of Electrical and Computer Engineering Johns Hopkins University Baltimore MD USA

ISBN: (纸本)9781479903573

How do humans attend to and pick out relevant auditory objects amongst all other sounds in the environment? Based on neurophysiological findings we propose two task oriented attentional mechanisms acting as Bayesian priors which act on two separate levels of processing: a sensory mapping stage and object representation stage. The former sensory stage is modeled as a high dimensional mapping which captures the spectrotemporal nuances and cues of auditory objects. The latter object representation stage then captures the statistical distribution of the different classes of acoustic scenes. This scheme shows a relative improvement in performance by 81 % compared to a baseline system.

关键词： Auditory Attention Acoustic Scene Analysis Sensory processing Object based attention Sensory Process statistical distribution capture

来源：评论

学校读者我要写书评

暂无评论

DUTH at TREC 2013 Contextual Suggestion Track 22

DUTH at TREC 2013 Contextual Suggestion Track

引用

22nd Text REtrieval Conference, TREC 2013

作者： Drosatos, George Stamatelatos, Giorgos Arampatzis, Avi Efraimidis, Pavlos S. Electrical & Computer Engineering Dept. Democritus University of Thrace University Campus Xanthi67 100 Greece Institute for Language & Speech Processing Athena Research & Innovation Center University Campus Xanthi67 100 Greece

In this report we give an overview of our participation in the TREC 2013 Contextual Suggestion Track. We present an approach for context processing that comprises a newly designed and fine-tuned POI (Point Of Interest) data collection technique, a crowdsourcing approach to speed up data collection and two radically different approaches for suggestion processing (a k-NN based and a Rocchio-like). In the context processing, we collect POIs from three popular place search engines, Google Places, Foursquare and Yelp. The collected POIs are enriched by adding snippets from the Google and Bing search engines using crowdsourcing techniques. In the suggestion processing, we propose two methods. The first submits each candidate place as a query to an index of a user's rated examples and scores it based on the top-k results. The second method is based on Rocchio's algorithm and uses the rated examples per user profile to generate a personal query which is then submitted to an index of all candidate places. The track evaluation shows that both approaches are working well;especially the Rocchio-like approach is the most promising since it scores almost firmly above the median system and achieves the best system result in almost half of the judged context-profile pairs. In the final TREC system rankings, we are the 2nd best group in MRR and TBG, and 3rd best group in P@5, out of 15 groups in the category we participated. © 2013 22nd Text REtrieval Conference, TREC 2013 - Proceedings. All Rights Reserved.

关键词： Crowdsourcing

来源：评论

学校读者我要写书评

暂无评论

Predictive analysis of two tone stream segregation via extended Kalman filter

Predictive analysis of two tone stream segregation via exten...

引用

Annual Conference on Information Sciences and Systems (CISS)

作者： Debmalya Chakrabarty Mounya Elhilali Department of Electrical and Computer Engineering Laboratory of Computational Audio Perception Center for Speech and Language Processing Johns Hopkins University MD USA

Hearing engages in a seemingly effortless way, complex processes that allow our brains to parse the acoustic environment around us into perceptual sound objects, in a phenomenon called streaming or stream segregation. In this paper, we explore the hypothesis that the auditory system relies on the regularity inherent to each stream to segregate it from other competing streams in the scene. Tracking these regularities is achieved via a recursive prediction that tracks the evolution of each stream, using a Kalman filtering approach. The proposed approach combines spectral analysis operating at the level of the auditory periphery with a temporal analysis using Kalman tracking. To incorporate nonlinear relationships in the signal patterns, we employ an extended Kalman filter. This scheme is tested on sinusoidal patterns, or the two tone paradigm. The combined spectral and temporal analysis developed here is able to predict perceptual results of stream segregation by human listeners in a two tone paradigm.

关键词： Indexes Kalman filters Frequency response

来源：评论

学校读者我要写书评

暂无评论

Phonetically balanced and psychometrically equivalent monosyllabic word lists for word recognition testing in Thai

引用

Proceedings of Meetings on Acoustics 2015年第1期25卷

作者： Sajeerat Poonyaban Pasinee Aungsakulchai Charturong Tantibundhit Chutamanee Onsuwan Rattinan Tiravanitchakul Krit Kosawat Adirek Munthuli 1Center of Excellence in Intelligent Informatics Speech and Language Technology and Service Innovation (CILS) Thammasat University Thailand 2Department of Electrical and Computer Engineering Faculty of Engineering Thammasat University Rangsit Campus Khlong Luang Pathum Thani Thailand pond_poonyaban@*** meilyp.aung@*** tchartur@engr.tu.ac.th ***@student.tu.ac.th 3Department of English and Linguistics Faculty of Liberral Arts Thammasat University Rangsit Campus Khlong Luang Pathum Thani Thailand consuwan@tu.ac.th 4Department of Communication Science and Disorders Faculty of Medicine Mahidol University Thailand rartv@*** 5NECTEC National Science and Technology Development Agency (NSTDA) Pathum Thani Thailand krit.kosawat@nectec.or.th

Word recognition testing may be defined as a procedure to assess a listener’s ability to identify one-syllable words (such as phonetically-balanced/PB words) that are presented at a given suprathreshold level to arrive at a word recognition score. For Thai, Thammasat University and Ramathibodi Hospital Phonetically Balanced Word Lists 2015 (TU-RAMA PB’15) were created with five lists, each with 25 monosyllabic words. Besides its phoneme distributions being based on large-scale Thai spoken corpora, TU-RAMA PB’15 is in line with TU PB’14 with emphasis on phonetic balance, symmetrical phoneme occurrence, and word familiarity. To evaluate its homogeneity in terms of decibel intelligibility, the lists were recorded and presented to 10 normal hearing participants, ranging from 0 to 50 dB HL in 2 dB increments (ascending order) until they repeated correct verbal responses. Using logistic regression, regression slopes and intercepts were calculated to estimate percentage of correct performance at any given intensity and to construct psychometric functions for every list. Derived psychometric function slopes ranged from 0.2015 to 0.2262 while intensities required for 50% intelligibility ranged from 17.0876 to 20.8856. Two-way Chi-Square analysis performed on both parameters indicated that there was no significant difference among the five lists.

关键词：

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：