Investigation into the Use of WFSTs and DNNs for Speech Activity Detection in Broadcast Data Transcription

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F46747885%3A24220%2F17%3A00004824" target="_blank" >RIV/46747885:24220/17:00004824 - isvavai.cz</a>

Výsledek na webu

<a href="http://dx.doi.org/10.1007/978-3-319-67876-4_16" target="_blank" >http://dx.doi.org/10.1007/978-3-319-67876-4_16</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/978-3-319-67876-4_16" target="_blank" >10.1007/978-3-319-67876-4_16</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Investigation into the Use of WFSTs and DNNs for Speech Activity Detection in Broadcast Data Transcription

Popis výsledku v původním jazyce

This paper deals with the task of Speech Activity Detection (SAD). The main goal is to investigate a new SAD approach suitable for offline as well as online transcription of various radio/TV broadcasts containing a large amount of non-speech segments. For this purpose, Deep Neural Networks (DNNs) with various hyper-parameters are adopted and evaluated. Their training is carried out using artificially created mixtures of speech and non-speech signals. Our SAD scheme also utilizes a decoder based on Weighted Finite State Transducers (WFSTs). The decoder smooths the output from DNN, can operate online and utilizes context-based transduction model, where both speech and non-speech events are modeled using sequences of states. The final evaluation of the developed approach is carried out on standardized QUT-NOISE-TIMIT data set for SAD and in a real broadcast transcription system. The obtained results show that our SAD module yields state-of-the-art results on QUT-NOISE-TIMIT, and, at the same time, it is capable of (a) operating with low latency and (b) reducing the computational demands and error rate of the transcription system.

Název v anglickém jazyce

Investigation into the Use of WFSTs and DNNs for Speech Activity Detection in Broadcast Data Transcription

Popis výsledku anglicky

This paper deals with the task of Speech Activity Detection (SAD). The main goal is to investigate a new SAD approach suitable for offline as well as online transcription of various radio/TV broadcasts containing a large amount of non-speech segments. For this purpose, Deep Neural Networks (DNNs) with various hyper-parameters are adopted and evaluated. Their training is carried out using artificially created mixtures of speech and non-speech signals. Our SAD scheme also utilizes a decoder based on Weighted Finite State Transducers (WFSTs). The decoder smooths the output from DNN, can operate online and utilizes context-based transduction model, where both speech and non-speech events are modeled using sequences of states. The final evaluation of the developed approach is carried out on standardized QUT-NOISE-TIMIT data set for SAD and in a real broadcast transcription system. The obtained results show that our SAD module yields state-of-the-art results on QUT-NOISE-TIMIT, and, at the same time, it is capable of (a) operating with low latency and (b) reducing the computational demands and error rate of the transcription system.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

20204 - Robotics and automatic control

Projekt

<a href="/cs/project/TA04010199" target="_blank" >TA04010199: MULTILINMEDIA - Multilinguální platforma pro monitoring a analýzu multimédií</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)<br>I - Institucionalni podpora na dlouhodoby koncepcni rozvoj vyzkumne organizace

Rok uplatnění

2017

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Communications in Computer and Information Science

ISBN

978-331967875-7

ISSN

1865-0929

e-ISSN

—

Počet stran výsledku

18

Strana od-do

341-358

Název nakladatele

Springer Verlag

Místo vydání

Spolková republika Německo

Místo konání akce

Lisbon; Portugal

Datum konání akce

1. 1. 2016

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—