Modeling and Integration of Prosody for Speaker, Language and Speech Recognition

Abstract

Methods for extraction and representation of prosodic features were discussed in the previous chapter. Now the extracted prosodic features should be represented in a manner useful for the specific recognition task. Since prosody reflects the speaker-specific, language-specific and sound-specific. Representation should be such a way that it brings out the differences among speaker/language/sound units. Then it should be modeled appropriately. Different modeling techniques have been employed by the researchers for building prosodic models. Various methods of integration have been employed for combining evidence from prosodic models with evidence from other knowledge sources such as acoustic models and language models.