Progress Report - V Ravi Chander ASR for the Elderly Progress Report - V Ravi Chander
Organisation of Talk Goal and Research Questions Experimental set-up Results Next Steps
Goal Project : MATCH (Mobilizing Advanced Technologies for Care at Home) Role : ASR for spoken dialogue system in home care environment for elderly people.
-different demographics Research Questions (1) Adaptation Reduce the amount of annotation required. Existing Recogniser -different domain -different demographics MATCH
Research Questions (2) Use corpus from a different domain to improve performance in the target domain.
Approach Only a small corpus of MATCH exists. Validate the ideas using AMI corpus.
Experimental Set-up AMI 1 AMI 2 AMI 3 AMI 4 CTS-NIST BASELINE Adapted ASR Test AMI 5
Sampling Techniques Random Supervised (likelihood based) Speaker level Utterance level Unsupervised GMM based closest speakers Utterances selected based on potential informativeness
GMM based closest Speakers No of speakers: 136 GMM for each speaker Feature vectors used : 13 MFCC coefficients Only speech frames used for modelling No of Gaussian components : 32
Distance measure KL divergence between Gaussians Distance between GMMs f g
Clustering (using CluTo)
Clustering Results Good pattern seen in age: If seen at the top two clusters: Females: 77.3% Males: 62.5% If seen at the level of 4 clusters (regrouping according to known information): Females: 92.45% Males: 98.79% No pattern seen with respect to age.
Next steps (1) Create one GMM from half of AMI 5 test speakers data (AMI 5). Sample the closest speakers from AMI 3 and AMI 4 to the above GMM, to adapt the baseline recognizer. Test against the other half. Repeat the experiment interchanging the two halves.
Next Steps (2) Utterance sampling based on the potential informativeness of the sample. Data Buffer Baseline ASR Adaptation corpus ranking sampling labelling
Thank you!! Questions / Suggestions