Idiap on LinkedIn Idiap youtube channel Idiap on Twitter Idiap on Facebook
Personal tools
You are here: Home Awards and Jobs News New opening for 4 Internship positions

New opening for 4 Internship positions

1.Uncertainty decoding and probabilistic features

The goal here is to use probabilistic features in ASR. These could come either from noisy observations, or from features that have some kind of probabilistic nature. The student would have to:

  • Extract features plus some associated variance
  • Modify Juicer to enable these to be processed
  • Compare against some known method for dealing with such features (e.g., noise robustness)

More information about the position can be found on our online recruitment system


2. Conversational TTS

The idea here is to train an HTS system using conversational (AMI meeting) data. The difficulty is in the annotation of the data to a level suitable for use in TTS training. It would involve:

  • Identification of suitably clean segments
  • Use of ASR to label the segments
  • Training of HTS system
  • Listening tests to evaluate the result

More information about the position can be found on our online recruitment system


3. Keyword spotting

Multimodal KWS

The idea is to create a KWS system which would be able to spot keywords independent of input language. It would have to be based on a phoneme recognition system trained multi-lingually (e.g., International Phoneme Alphabet, ...). A difficulty would be to come up with some lexicon (which might be rule based, e.g., G2P). Comparison would be made with a KWS created just for a unique language.

Online KWS

The goal is to develop a KWS which would be able to perform real-time. The work might not be so comlex and could be based either on tracter features or directly on BSAPI library based features.

Then phoneme based KWS would be used, or we could hack an SVite decoder to be able to produce lattices in real-time (e.g., over short segments) and then simply search the lattice output.

More information about the position can be found on our online recruitment system


4. Subspace Gaussian Mixture modelling (SGMM) for multi-lingual ASR

The goal is to continue a research direction based on a new acoustic modeling paradigm which has also a relationship with the popular Joint Factor Analysis dominating speaker identification systems.

The point would be to repeat a base-line system presented at ICASSP 2010 and then potentially continue the research with a new ASR toolkit (KALDI) which was developed for such new acoustic models.

More information about the position can be found on our online recruitment system

Sep 20, 2010
Document Actions