Implementation and analysis of deep learning models for speech classification
summary
Data set is assembled from recordings and TV audio into 25 hours, 5,000 code-switched samples and 2,000 English plus 2,000 Hindi 10s clips. Pipeline extracts Mel spectrograms and MFCCs with librosa, trains a CNN on Mel spectrograms for image-style classification and an LSTM RNN on MFCC + LSTM for temporal modeling. Spectrograms underperform; Mel spectrograms give clearer heat maps, both models are trained and evaluated to label Hindi, English, or code-switched audio.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:09Agenda and problem
- 3:15Code switching in speech
- 5:43Data collection and preprocessing
- 8:34Convolutional Neural Network pipeline
- 10:24Recurrent Neural Network pipeline
- 12:12Summary
- 14:20Question and answer session
related talks
28mHow to use Deep Learning to process audio data and classify sounds.
40mGathering Insights from Audio Data
25mLooking at Sound
31mClassifying Audio into Types using Python
45mAuditory Intelligence : beyond speech recognition
22mDeep learning module recognizes video sounds
11mClassifying Spoken Location Names
40mBuilding Features and Machine Learning Models for Audio Signals