| Title |
Database Adaptation for Speech Recognition in Cross-Environmental |
| Authors |
Oren Gedge (NSC- Natural Speech Communication, Scansoft, DaimlerChrysler AG) Christophe Couvreur ( NSC- Natural Speech Communication, Scansoft, DaimlerChrysler AG) Klaus Linhard ( NSC- Natural Speech Communication, Scansoft, DaimlerChrysler AG) Shaunie Shammass (NSC- Natural Speech Communication, Scansoft, DaimlerChrysler AG) Ami Moyal (NSC- Natural Speech Communication, Scansoft, DaimlerChrysler AG) |
| Session |
SO5: Speech Variabilities & Multilingual ASR |
| Abstract |
This study aims to simulate conditions that reflect the needs of speech-controlled consumer devices. In particular, it must be ascertained whether training in one type of environmental condition can be effectively adapted to other acoustic conditions, without having to perform costly collection in each specific type of environment. The adaptation tool performs two tasks: convolution of the clean speech signal with a given (room) Impulse Response (IR) and addition of noise to the convolved speech signal. Noise addition is done using recordings of typical environmental noise sources. Baseline, cross-tests and adaptation tests were performed. Results of the convolution and noise addition tests are presented for a speaker-dependent name recognition task. It is shown that adaptation reduces the recognition error rates when compared to the cross-tests. Ongoing tests within the SPEECON project are currently underway for evaluating the effectiveness of straight noise addition after convolution. For the speaker-independent case, preliminary tests on a database specifically collected for testing purposes have been performed. |
| Keywords |
Adaptation, Speech recognition, Acoustic environment |
| Full Paper |