Department of Computer Graphics and Multimedia
Exploiting Language Information for Situational Awareness (ELISA)
|Czech title:||Využití jazykových informací pro informování v různých situacích (ELISA)|
|Reseach leader:||Burget Lukáš|
|Team leaders:||Černocký Jan, Matějka Pavel, Szőke Igor|
|Team members:||Glembek Ondřej, Ondel Lucas|
|Agency:||University of Southern California|
|Keywords:||Speech processing, language, apeech mining|
|Speech processing in our proposal will be addressed by low-resource or language-agnostic technologies. Rather than concentrating on mining the content (for which, obviously, standard resources such as acoustic model, language model or pronunciation dictionary will be lacking), speech data will be handled by a multitude of "speech miners" that make minimum use of resources of the target language.
The processing will begin with a reliable voice activity detection (VAD) capable of segmenting the signal into useful and useless portions. Often regarded as "not a rocket science", a good VAD is crucial for correct functioning of the following blocks and for human processing of speech input. Our work will improve on existing DNN-based VAD that proved its efficiency in a difficult RATS setting [Ng2012]. A processing with several phone posterior estimators with either mono-lingual or multilingual phoneme sets [Schwarz2009] will follow to provide the "miners" with a coherent low-dimensional representation.
The first real "miner" will be language identification (LID) with a significant set of target languages (>60). Even if it is not sure that the target language will be in this set, LID will allow to detect segments in English or possibly in other languages for which we have ASR technology. We will follow our recent development of LID base on features derived from phone posteriors [Plchot2013] as well as on DNNs. We will also work on enrollment of a new language with very little data (down to one utterance). Another "miner" will perform basic speaking style recognition allowing to separate read speech from spontaneous. Finally, speaker recognition (SRE) or clustering will allow to gather information about speakers (in case they were previously enrolled) or at least to perform coarse speaker clustering, as for the analyst, the information on who is speaking can be equally important as what is said. Here, we will build up on our significant track in iVector-based SRE and will mainly work on automatic adaptation and calibration on unlabeled data-sets [Brummer2014]|