NVIDIA CUDA application to train and decode the Hidden Markov Models
The urgency of the discussed issue is caused by the need of optimization of huge speech corpus's processing algorithms required for developing robust automatic speech recognition systems. The evolution of modern multicore processors, specifically graphical processor units GPU, allows improving...
Main Authors: | , |
---|---|
Format: | Article |
Language: | Russian |
Published: |
Tomsk Polytechnic University
2019-05-01
|
Series: | Известия Томского политехнического университета: Инжиниринг георесурсов |
Subjects: | |
Online Access: | http://izvestiya.tpu.ru/archive/article/view/1374 |
_version_ | 1797813203434995712 |
---|---|
author | Pavel Zatsepin Denis Gefke |
author_facet | Pavel Zatsepin Denis Gefke |
author_sort | Pavel Zatsepin |
collection | DOAJ |
description | The urgency of the discussed issue is caused by the need of optimization of huge speech corpus's processing algorithms required for developing robust automatic speech recognition systems. The evolution of modern multicore processors, specifically graphical processor units GPU, allows improving sufficiently the performance of difficult and resource-intensive digital signal processing algorithms and reducing sufficiently a data processing time. The main aim of the study is to optimize education (Baum-Welch re-estimation) and decoding (Viterbi) algorithms of Hidden Markov Models by parallel programming technology NVIDIA CUDA and to estimate performance increase in comparison within the CPU. The methods used in the study: the search of education and decoding algorithm's parts suitable for effective parallel realization by NVIDIA CUDA and its implementation. The results: The authors have developed parallel realization of education and decoding Hidden Markov Models algorithms by GPU and have estimated the performance increase in comparison within the CPU for different model's parameters (the number of model state and dimension of a feature vector). The results of the paper can be used both by engineers developing and improving the automatic speech recognition systems and by explorers working on a digital signal processing and artificial intelligence systems. |
first_indexed | 2024-03-13T07:48:48Z |
format | Article |
id | doaj.art-37e822f316e5480a978b5d7cf479c17a |
institution | Directory Open Access Journal |
issn | 2500-1019 2413-1830 |
language | Russian |
last_indexed | 2024-03-13T07:48:48Z |
publishDate | 2019-05-01 |
publisher | Tomsk Polytechnic University |
record_format | Article |
series | Известия Томского политехнического университета: Инжиниринг георесурсов |
spelling | doaj.art-37e822f316e5480a978b5d7cf479c17a2023-06-02T21:08:24ZrusTomsk Polytechnic UniversityИзвестия Томского политехнического университета: Инжиниринг георесурсов2500-10192413-18302019-05-013245NVIDIA CUDA application to train and decode the Hidden Markov ModelsPavel ZatsepinDenis GefkeThe urgency of the discussed issue is caused by the need of optimization of huge speech corpus's processing algorithms required for developing robust automatic speech recognition systems. The evolution of modern multicore processors, specifically graphical processor units GPU, allows improving sufficiently the performance of difficult and resource-intensive digital signal processing algorithms and reducing sufficiently a data processing time. The main aim of the study is to optimize education (Baum-Welch re-estimation) and decoding (Viterbi) algorithms of Hidden Markov Models by parallel programming technology NVIDIA CUDA and to estimate performance increase in comparison within the CPU. The methods used in the study: the search of education and decoding algorithm's parts suitable for effective parallel realization by NVIDIA CUDA and its implementation. The results: The authors have developed parallel realization of education and decoding Hidden Markov Models algorithms by GPU and have estimated the performance increase in comparison within the CPU for different model's parameters (the number of model state and dimension of a feature vector). The results of the paper can be used both by engineers developing and improving the automatic speech recognition systems and by explorers working on a digital signal processing and artificial intelligence systems.http://izvestiya.tpu.ru/archive/article/view/1374speech recognitionparallel computingHidden Markov ModelsNVIDIA CUDAViterbi algorithmBaum-Welch re-estimation algorithm |
spellingShingle | Pavel Zatsepin Denis Gefke NVIDIA CUDA application to train and decode the Hidden Markov Models Известия Томского политехнического университета: Инжиниринг георесурсов speech recognition parallel computing Hidden Markov Models NVIDIA CUDA Viterbi algorithm Baum-Welch re-estimation algorithm |
title | NVIDIA CUDA application to train and decode the Hidden Markov Models |
title_full | NVIDIA CUDA application to train and decode the Hidden Markov Models |
title_fullStr | NVIDIA CUDA application to train and decode the Hidden Markov Models |
title_full_unstemmed | NVIDIA CUDA application to train and decode the Hidden Markov Models |
title_short | NVIDIA CUDA application to train and decode the Hidden Markov Models |
title_sort | nvidia cuda application to train and decode the hidden markov models |
topic | speech recognition parallel computing Hidden Markov Models NVIDIA CUDA Viterbi algorithm Baum-Welch re-estimation algorithm |
url | http://izvestiya.tpu.ru/archive/article/view/1374 |
work_keys_str_mv | AT pavelzatsepin nvidiacudaapplicationtotrainanddecodethehiddenmarkovmodels AT denisgefke nvidiacudaapplicationtotrainanddecodethehiddenmarkovmodels |