Automatic speech recognition of Venezuelan Spanish: Test results on phrases pronounced by women

• José Luciano Maldonado Universidad de Los Andes-Venezuela

Abstract

The objective of this work was to test the automatic recognition of phrases as pronounced by Venezuelan women. The activities included signal processing for training and testing, selection and construction of voice models, construction of the recognizer and recognition tests. The sounds were modeled using Hidden Markov Models and HTK, a publicly available toolkit. The sound files are a subset of the Venezuelan SpeechDat database. Tests were done on dates. The capacity of recognition was observed at the following levels: phonemes, words and entire phrases for each test file. To determine if the sentences were appropriate, the recognizer was equipped with a grammar which included most acceptable ways of saying dates according to Venezuelan usage. The ranges of best results were the following: at a word level, between 89.14% and 97.01%, and for entire phrases between 38.24% and 71.52%. This effort at automatic speech recognition independent of the speaker with Venezuelan voices lets us state that in the future recognizers can be built which will be able to handle the voice characteristics corresponding to this country”™s speech.