Looking at Sound
summary
Raw audio becomes images by FFT into mel spectrograms (24 kHz, librosa) plus delta channels, fed to CNNs with inception‑style modules, frequency‑band splits, and time‑only max‑pooling; between‑class mixing, pitch shifts and SpecAugment reduce overfitting (dropout removed) and raise accuracy by a few percent on UrbanSound8K where humans hit ~82% on 4s clips. Event segmentation uses simple peak detection; emotion work layers speech‑to‑text and word embeddings into the same pipeline.
AI-written summary · about
chapters (13)
- 0:00Sound classification project
- 1:25Urban sound challenges
- 3:18Raw audio and Fourier basics
- 5:26Mel spectrograms from audio
- 7:14Feature engineering with delta
- 9:01Between-class learning for augmentation
- 10:55Advanced audio augmentation
- 13:13Network architecture for sound
- 16:48Inception and ResNet audio models
- 18:17Frequency bands for specialized networks
- 19:42Full audio event detection pipeline
- 20:45Key takeaways and performance
- 21:06Emotion detection in contact centers
related talks
28mHow to use Deep Learning to process audio data and classify sounds.
22mDeep learning module recognizes video sounds
45mAuditory Intelligence : beyond speech recognition
40mGathering Insights from Audio Data
01h 14mExploring the hidden potential of sound data
59mTechnical Workshop: Sarcastically Speaking: Unlocking Multi-modal Sentiment Analysis With NLP and Facial Expressions
22mImplementation and analysis of deep learning models for speech classification
40mBuilding Features and Machine Learning Models for Audio Signals