Cross-language Transfer of Multilingual Phoneme Models

We present a method to use speech data from multiple languages to enhance the performance of a flexible vocabulary command word recognizer which is trained using a small amount of speech data of the target language. We develop data-driven approaches for identification of multilingual phoneme units and mapping of these units to the target language phonemes, and evaluate them against the knowledge based approach of mapping identical SAMPA phoneme symbols. The usefulness of multilingual context dependent phoneme modeling for cross-language transfer is shown. Our method achieves significant improvement of recognition performance in the target languages Danish and English by cross-language transfer of multilingual models trained on French, German, Italian, Portuguese and Spanish speech if phonetically rich target language speech data by less than 100 speakers of roughly 1/2 minute duration per speaker is available.