Direction-of-Voice (DoV) Estimation for Intuitive Speech Interaction with Smart Device Ecosystems
summary
System builds a lightweight DoV estimator using a four-channel mic array, physics-informed features (autocorrelation, SRMR reverberation measure, GCC-PHAT cross-correlations, FFT-derived high/low power ratios and polynomial coefficients) and trains an Extra Trees Classifier on 11,520 recordings (350+ minutes) across rooms, distances (1/3/5 m) and utterances. Accuracy is 97% at 1 m, 93% at 3 m and 88% at 5 m (about 90% overall, comparable to human baseline); it does not handle multiple simultaneous speakers and falls short of 99% consumer expectations.
AI-written summary · about
chapters (14)
- 0:00Intro
- 1:10Smart Voice Interfaces
- 2:43"Direction of Voice" Estimation
- 3:27Hardware
- 4:07Crispness of First Wavefront
- 4:51Voice Frequency Distribution
- 6:08Human Benchmarking
- 7:35Definitions of "Facing"
- 7:54Data Collection
- 8:50ML Stack
- 9:10Evaluation
- 9:47Cross utterance
- 10:55Across Speaker Distances
- 11:19Limitations and Future Work
related talks
06mSensing to Hear: Speech Enhancement for Mobile Devices Using Acoustic Signals
13mVOICE Summit 2019 | VED Talks | Jim Harper | July 24
06mToward User-Driven Sound Recognizer Personalization with People Who Are d/Deaf or Hard of Hearing
14mVeCare: Statistical Acoustic Sensing for Automotive In-Cabin Monitoring
22mUbicoustics: Plug-and-Play Acoustic Activity Recognition
13mVOICE Summit 2019 | VED Talks | Alexandros Potamianos | July 23
25mBeyond Transcription: Building Voice AI That Understands Conversations
20mVoice fingerprinting & recognition