Connectionists: Master2 (research) position at Multispeech Team, LORIA and INRS (Nancy, France)

[Irina Illina]

Master2 (research) position at Multispeech Team, LORIA and INRS (Nancy, France) 

Speech intelligibility: how to determine 

the degree of nuisance 

Irina Illina, Patrick Chevret 




General information 

Supervisors 

    * 

Irina Illina , LORIA, Campus Scientifique - BP 239, 54506 Vandoeuvre-lès-Nancy, 



Tél : 03 54 95 84 90, bureau C147, illina at loria.fr 




    * 

Patrick Chevret, INRS, 1 rue du Morvan, 54519 Vandoeuvre-lès-Nancy, 



Tél : 03 83 50 85 67, patrick.chevret at inrs.fr 







Motivations 

The intelligibility of speech means the ability of a conversation to be understood by a listener located nearby. The level of speech intelligibility depends on several criteria: the level of ambient noise, the possible absorption of part of the sound spectrum, acoustic distortion, echoes, etc. The intelligibility of speech is used to assess the performance of telecommunication systems or absorption in rooms. 




The speech intelligibility can be evaluated: 

    * 

subjectively: listeners hear several words or sentences and answer different questions (the transcription of sounds, the percentage of perceived consonants, etc.). The scores are the value of intelligibility ; 
    * 

objectively, without involving listeners, using acoustic measures: the index of speech intelligibility (speech transmission index, STI) and the interference level with speech. 





Subjective measures are dependent of listeners and require a large number of listeners. This is difficult to achieve, especially when there are different types of environments. Moreover, it is necessary to evaluate this measure for each listener. Objective measures have the advantage of being automatically quantifiable and to be precise. However, which objective measures can measure the nuisance of the environment on the intelligibility of speech and people's health remains an open problem. For example, the STI index consists of measuring the energy modulation. But the energy modulation can be produced by the machines, yet it does not match the speech. 




Subject 

In this internship, we focus on the study of various objective measures of speech intelligibility. The goal is to find reliable measures to evaluate the level of nuisance of environment to speech understanding, to long-term mental health of people and to productivity. Some possible solutions consist to correlate the word confidence measure, noise measurement confidence and subjective measures of speech intelligibility. To develop these measures, the automatic speech recognition system will be used. 




This internship will be performed through collaboration between our Multispeech team of LORIA and INRS ( National Institute of Research and Safety ). INRS works on professional risk identification, analysis of their impact on health and prevention. INRS has a rich corpus of recordings and subjective measures of speech intelligibility. This corpus will be used in the context of this internship. Our Multispeech team has great expertise in signal processing and has developed several methodologies for noise estimation. The Multispeech team developed the complete system of automatic speech recognition. 







Required skills 

Background in statistics and object-oriented programming. 




References 

[1] S. Wijngaarden, J. Verhave, H. Steeneken The Speech Transmission Index after four decades of development, 2012 

[2] P. W. Barnett Overview of speech intelligibility Proc. I.O.A Vol 21, Part 5, 1999. 





  

  
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20151027/878bad9c/attachment.html>