8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

Speech Recognition Experiments with the SPEECON Database using Several Robust Front-Ends

Pere Pujol, Jaume Padrell, Climent Nadeu, Dusan Macho

Universitat Politecnica de Catalunya (UPC), Spain

In this paper we deal with the robustness problem in speech recognition, using a Spanish subset of the recently collected SPEECON database, and focusing on the front-end side of the recognizer. Cross-microphone and cross-environment recognition tests are presented using both read and spontaneous continuous speech utterances. Our semi-continuous sub-word HMM back-end was fixed for all the tests. For comparison, we used both the clean-speech and the noisy-speech cepstrum-based ETSI standard front-ends, as well as a few relatively simple variants of the front-end that is based on frequency-filtering (FF) features. In all our tests, the best word error rates scores were obtained with the FF front-end. Moreover, a technique based on a long-term log spectral mean subtraction was successfully used to reduce the reverberation affecting the utterances from the furthest microphones.

Full Paper

Bibliographic reference.  Pujol, Pere / Padrell, Jaume / Nadeu, Climent / Macho, Dusan (2004): "Speech recognition experiments with the SPEECON database using several robust front-ends", In INTERSPEECH-2004, 2105-2108.