DNN-HMM Acoustic Modeling for Large Vocabulary Telugu Speech Recognition

Vegesna, Vishnu Vidyadhara Raju; Gurugubelli, Krishna; Vydana, Hari Krishna; Pulugandla, Bhargav; Shrivastava, Manish; Vuppala, Anil Kumar

doi:10.1007/978-3-319-71928-3_19

Cited by 9 publications

(5 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A basic system of hybrid speech recognition based on DNN-HMM acoustic model is established in the literature [11]. At the same time, according to the advantages of TDNN and LSTM network in time series dependent information modeling, TDNN-LSTM network is introduced into the acoustic model of mixed language speech recognition to further reduce the word error rate [12]. The literature used multi-objective task learning method to train the model.…”

Section: Related Workmentioning

confidence: 99%

Research on language simulation and speech recognition based on data simulation of Machine Learning System

Qin-ke¹,

Honggai

2023

Preprint

View full text Add to dashboard Cite

In recent years, human-computer interaction project has become a mainstream research topic in the field of artificial intelligence. Among them, the most important interaction mode is voice interaction mode, which is based on voice recognition technology and is also the main factor to promote the development of artificial intelligence technology itself. In addition, speech recognition and modeling in noisy background have also been fully developed. In the actual communication environment, in addition to the voice information of the communicator, there are also noisy background sounds, which will reduce the accuracy of speech recognition. Therefore, it is necessary to conduct modeling research on this problem to improve the recognition ability of the speech recognition model. In this context, this paper combines machine learning technology and artificial intelligence technology to conduct in-depth research to improve the performance of speech recognition model, achieve the balance between efficiency and performance, and run stably and easily. The system is mainly divided into user registration, data flow input and processing, identification model and output results and other modules. The experimental data show that the response time of the system is proportional to the number of test samples. Therefore, as long as resources are allocated reasonably, stable system response time can be obtained, and system performance and scalability can fully meet the requirements of language simulation tasks. In this paper, the optimization of language simulation system is completed through the comprehensive study of machine learning and artificial intelligence technology.

show abstract

Section: Related Workmentioning

confidence: 99%

Research on language simulation and speech recognition based on data simulation of Machine Learning System

Qin-ke¹,

Honggai

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Acoustic Modeling plays a vital role in speech recognition systems. The statistical methods like HMM and GMM are the most popular choice for developing ASR system [23]. Huge research has been done to increase the evaluation speed of GMM & to optimize the tradeoff between their flexibility & the amount of training data to avoid the serious overfitting [24].…”

Section: Literature Review and Previous Work Comparisonmentioning

confidence: 99%

“…MFCC ) tandem or bottleneck features generated using MM [25,26]. As an alternative of conventional HMM & GMM-HMM [23],SGMM & subspace GMM [27], DNN [28,29,30] convolutional MM [31,32,33,30], RNN [34,35] has been proposed and matter of deep investigation. To solve Hindi LVCSR Problem various DNM acoustic modeling should be tested and some novel acoustic modeling should be proposed.…”

Section: Literature Review and Previous Work Comparisonmentioning

confidence: 99%

Automatic Hindi Speech Recognition: Challenges and Future Scop

Kumar¹,

Gupta²,

Tripathi³

et al. 2023

Artificial Intelligence and Communication Technologies

View full text Add to dashboard Cite

Automatic Speech Recognition (ASR) field is the witness of remarkable improvement over the last few years. Deep learning architectures are the main reason for ASR evaluation. Deep learning techniques tremendously improve ASR performance, but these architectures are data-hungry. There are more than 5000 languages in India that do not have such a huge amount of resources. Hindi is one of them. The lack of freely available Large Vocabulary Hindi speech datasets is the major hurdle for developing Hindi ASR. In this paper, we investigate the various work available for Hindi ASR. The available resource is also discussed with issues and challenges in Hindi ASR. This paper provides a detailed analysis of various techniques available in ASR with comparison to Hindi ASR. The future scope of Hindi ASR is also presented.

show abstract

“…Finally, the DNN are trained the estimation of the posterior probabilities of HMM states from given observation sequences. The weight of DNN training is done by minimizing the cost function of cross entropy criterion [19].…”

Section: Dnn-hmm Based Acoustic Modelmentioning

confidence: 99%

Discrete Wavelet Denoising into MFCC for Noise Suppressive in Automatic Speech Recognition System

Naing¹,

Hidayat²,

Hartanto³

et al. 2020

IJIES

View full text Add to dashboard Cite

Automatic Speech Recognition (ASR) is a challenging task and the most problematic issues being in presence of background noise and substantial variability in speech. Extracting the noise-robust features adjust for speech degradations due to noise effect retained popular issue in recent years. This paper presented a framework for wavelet denoising scheme and analysed the different wavelet families and proper thresholding rule into feature extraction to enhance the performance of ASR system. Gaussian Mixture Model-based Hidden Markov Model (GMM-HMM) and Deep Neural Network (DNN)-HMM are used as the speech recognizer. The recognition performance shows that the noise-robust features are obtained while combining with the wavelet transform denoising into Mel Frequency Cepstral Coefficient (MFCC) on Aurora2 database. The best accuracy is gained by cross entropy DNN-HMM training using denoising with Coiflet wavelet and Rigrsure threshold, which provides 97.54% in 10dB, 93.13% in 5dB, 75.63% in 0dB and 37.29% in −5dB.

show abstract

DNN-HMM Acoustic Modeling for Large Vocabulary Telugu Speech Recognition

Cited by 9 publications

References 15 publications

Research on language simulation and speech recognition based on data simulation of Machine Learning System

Research on language simulation and speech recognition based on data simulation of Machine Learning System

Automatic Hindi Speech Recognition: Challenges and Future Scop

Discrete Wavelet Denoising into MFCC for Noise Suppressive in Automatic Speech Recognition System

Contact Info

Product

Resources

About