Determination of Room Impulse Response for synthetic data acquisition and ASR testing

摘要

Automatic Speech Recognition (ASR) works best when the speech signal best matches the ones used for training. Training, however, may require thousands of hours of speech, and it is impractical to directly acquire them in a realistic scenario. Some improvement can be obtained by incorporating typical building acoustics measurement parameters such as RT, Cx, LF, etc., with limited gain. Instead, we estimate Room Impulse Responses (RIRs), and convolve speech and noise signals with the estimated RIRs. This produces realistic signals, which can then be processed by the audio pipeline, and used for ASR training. In our research, we use rooms with variable acoustics and repeatable source-receiver positions. The receivers are microphone arrays making the relative phase and magnitude critical. A standard mouth simulator for voice signals at various positions in the room is under robot control. A limited corpus of speech data as well as noise sources is recorded and the RIR at these 27 positions is determined using a variety of methods (chirp, MLS, impulse, and noise). The convolved RIR with the “clean speech” is compared to the actual measurements. Test methods used, differences from the measurements, and the difficulty of determining the unique RIR will be presented.