Automatic Pronunciation Evaluation And Mispronunciation Detection Using CMUSphinx

摘要

Feedback on pronunciation is vital for spoken language teaching. Automatic pronunciation evaluation and feedback can help non-native speakers to identify their errors, learn sounds and vocabulary, and improve their pronunciation performance. These evaluations commonly rely on automatic speech recognition, which could be performed using Sphinx trained on a database of native exemplar pronunciation and non-native examples of frequent mistakes. Adaptation techniques using target users' enrollment data would yield much better recognition of non-native speech. Pronunciation scores can be calculated for each phoneme, word, and phrase by means of Hidden Markov Model alignment with the phonemes of the expected text. In addition to the basic acoustic alignment scores, we have also adopted the edit distance based criterion to compare the scores of the spoken phrase with those of models for various mispronunciations and alternative correct pronunciations. These scores may be augmented with factors such as expected duration and relative pitch to achieve more accurate agreement with expert phoneticians' average manual subjective pronunciation scores. Such a system is built and documented using the CMU Sphinx3 system and an Adobe Flash microphone recording, HTML/JavaScript, and rtmplite/Python user interface.