Auditory Intelligence : beyond speech recognition
summary
Proposes concrete systems: an unsupervised speech model that jointly learns subword units and a dictionary and outperforms phoneme-based systems, finding 384 learned units as optimal; an audio-event pipeline built from a self-made dataset of 5,000 clips / 768 minutes using large-input-field + VGG-style deep CNN + data augmentation that yields state-of-the-art event detection and powers audiovisual highlight detection via deep audio features; and a music source-separator called MDenseNet with multi-band modules that wins major separation contests and produces convincing demos.
AI-written summary · about
chapters (28)
- 0:00Introduction
- 1:58Research focus
- 2:38Deep learning
- 4:01Problems
- 7:21Problem
- 8:20Why sound
- 9:51Outline
- 10:33Speech Recognition
- 13:33Semisupervised Deep Neural Networks
- 15:52Acoustic Model and Dictionary
- 18:16Conclusions
- 18:50Audio Event Recognition
- 19:27Data Collection
- 19:51Automatic Event Recognition
- 21:20Architecture
- 23:00A Version
- 24:49Audio Detection
- 25:22Audio Feature
- 26:04Deep Feature
- 26:41Demo
- 29:08Analysis
- 30:30Conclusion
- 30:53Audio Separation
- 33:18European Network for Audio
- 35:02Separation Contest
- 36:25Demonstration
- 37:50Sony
- 39:02Future of AI
related talks
25mLooking at Sound
22mDeep learning module recognizes video sounds
22mUbicoustics: Plug-and-Play Acoustic Activity Recognition
28mHow to use Deep Learning to process audio data and classify sounds.
06mSensing to Hear: Speech Enhancement for Mobile Devices Using Acoustic Signals
22mImplementation and analysis of deep learning models for speech classification
30mStarting deep learning projects without enough labelled data
01h 14mExploring the hidden potential of sound data