Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Rejection Techniques Based on Context Independent Subword Units

Juan Carlos Torrecilla, Daniel Tapias, F. Javier Caminero-Gil, Luis Villarrubia

Speech Technology Group Telefonica Investigation y Desarrollo, Madrid, Spain

This paper addresses the problem of how to identify and reject Out-Of-Vocabulary (00V) words and background noises using an isolated word speaker independent Hidden Markov Model (HMM) recognizer. The proposed method builds a garbage model as a combination of context independent subword HMM units to represent all possible 00V words and background noises. The rejection rate is adjusted by an appropriate penalty that depends on the utterance duration and then is applied to the garbage model acoustic likelihood. This approach makes the behavior of the system independent of the utterance duration. Additionally, we provide linguistic information to the garbage model so that illegal phone sequences are not allowed. This approach reduces the average branching factor of the garbage model and increases the rejection of the OOV words.

Full Paper

Bibliographic reference.  Torrecilla, Juan Carlos / Tapias, Daniel / Caminero-Gil, F. Javier / Villarrubia, Luis (1995): "Rejection techniques based on context independent subword units", In EUROSPEECH-1995, 1633-1636.