Using child directed speech to bootstrap a model based speech segmentation and recognition system

摘要

Method for acquiring a pool (P) of speech syllable models comprising the steps:
- finding a training segment (X) using unsuper-vised speech segmentation or SU spotting;
- if the model pool (P) is empty, training a first speech syllable model and adding it to the pool;
- else
o determining an existing model (F) in the model pool (P) that best matches the given segment (X);
o using the model (F) to score the new training segment (X);
o if the score is smaller than a predefined threshold, creating a new model for the training segment (X) and adding it to the pool; else using the new training segment (X) to improve or to re-estimate the model (F);
- repeating the above steps.