Research / Robust speech recognition and spoken search / [17]
CENATAV at ALBAYZIN 2018: a Kaldi system for finding spoken terms
Cenatav Voice Group System for Albayzin 2018 Search on Speech Evaluation. IberSPEECH 2018, 2018.
The CENATAV Voice Group's first entry in the ALBAYZIN evaluations: a speech recogniser and a term detector built with Kaldi. Of three acoustic models, the DNN-HMM gave the best-calibrated scores and was submitted as the primary system.
The task
In spoken term detection a list of written terms has to be found in audio files. The list is only known after the audio has been processed, so some terms are outside the recogniser’s vocabulary and need separate treatment.
The system
Two modules, both built with the open-source Kaldi toolkit. A large-vocabulary speech recogniser produces word lattices and alignments, and a detection module searches them for the terms. The recogniser was trained on the audio supplied by the organisers and on data from ELDA, with a lexicon and a trigram language model taken from the transcripts.
What we found
Three acoustic models were compared: GMM-HMM, S-GMM and DNN-HMM. On the MAVIR development set they scored an ATWV of 0.56, 0.59 and 0.59. The DNN-HMM model missed the fewest terms and showed the smallest gap between its maximum and actual scores, a sign of well-calibrated detection scores, so it was submitted as the primary system and the other two as contrastive systems.
Why it matters
It was the group’s first participation in the ALBAYZIN evaluations. The results of all participants are analysed in the ALBAYZIN 2018 evaluation paper.
Cite
Montalvo A.R., Ramírez J.M., Roble A., Calvo J.R. Cenatav Voice Group System for Albayzin 2018 Search on Speech Evaluation. IberSPEECH 2018, 2018.