Room Acoustic Characterization with Smartphone-based Automated Speech Recognition

Characterizing and monitoring the acoustic quality of a room is important for maintaining effective speech communication. Noise and echoes make speech harder to perceive, especially for individuals with auditory disabilities or remote conference participants. Reverberation time and noise level are common performance measures, but these are static measures and can be difficult for non-specialists to measure and interpret. In this work we consider smartphone-based automated speech recognition (ASR) as a proxy for speech intelligibility to assess room acoustics. We evaluate the error rate of the on-device ASR transcription of recordings of real speech and noise in a reverberant conference room. We also model the spatial processing benefits of binaural hearing by comparing recordings from an omnidirectional microphone (remote participant) and a microphone array (local participant). In a quiet room with a nearby microphone the ASR systems achieve near-perfect recognition rates, and performance degrades gradually as noise level or microphone distance are increased. Array processing improves robustness, providing an effective SNR improvement of around 3 dB. The results demonstrate that smartphone-based ASR can provide a convenient and readily available real-time assessment of the effects of noise and reverberation, even when those distortions are not apparent to local listeners.