DeepRange: Acoustic Ranging via Deep Learning
summary
A deep-learning acoustic ranging system uses a realistic simulator that injects white noise, multiple reflections, residual self-interference and randomized frequency response to create training data and trains a long-filter CNN that uses DRNet with kernel size 1440 and four channels. Evaluation runs on 160K test samples from 11 smartphones across 4 environments and 10 users. DRNet plus an ensemble reduces median distance error by about 3× at low SNRs compared to prior signal-processing methods.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:16Ranging based Applications
- 0:35Acoustic Ranging on Mobile
- 1:15Signal Processing Based Solutions
- 1:50Why is this interesting?
- 2:26How to enable data-driven approach?
- 3:02Realistic Simulation
- 3:29Frequency Response
- 3:59DNN Evolution
- 4:33Evaluation on real Trace ..163 K test samples
- 4:51Evaluation on SNR
- 5:12Ablation Study
- 5:46Contribution
related talks
45mThe Deep Learning Revolution in Automatic Speech Recognition
28mHow to use Deep Learning to process audio data and classify sounds.
12mDirection-of-Voice (DoV) Estimation for Intuitive Speech Interaction with Smart Device Ecosystems
56mRandom Embeddings, Matrix-valued Kernels and Deep Learning
25mFast deep recurrent net training
17mSLNet: A Spectrogram Learning Neural Network for Deep Wireless Sensing
22mUbicoustics: Plug-and-Play Acoustic Activity Recognition
37mDeep Recommender Systems