Explore multilingual speech representations for low-resource speech processing in this 40-minute talk by Brian Kingsbury from IBM. Discover how to achieve good automatic speech recognition performance with limited data for thousands of languages worldwide. Learn about the IARPA Babel Program, keyword search techniques, and various neural network architectures including Deep Neural Networks, Convolutional Neural Networks, and Recurrent Neural Networks. Understand the challenges and solutions for processing languages with limited resources, and gain insights into the importance of multilingual features in reducing the amount of data needed for training speech recognition systems in new languages. Examine three use cases and learn how more expressive architectures significantly impact performance. Ideal for researchers and professionals interested in advancing speech processing technologies for low-resource languages.
Multilingual Representations for Low-Resource Speech Processing
Overview
Syllabus
Intro
Why Care About Low-Resource Speech Processing?
How Much Transcribed Audio Do We Need?
Why Do We Need All That Training Data?
Multilingual Features
The IARPA Babel Program
Babel Languages
Limited resources
What is keyword search, and why focus on it?
How do we measure keyword search performance?
Properties of term-weighted value
Take-Home Messages
Three Ways of Looking at Speech
Deep Neural Network
A Stacked DNN Architecture
Convolutional Neural Network
Considered 2 CNN Architectures
Recurrent Neural Network
Bidirectional LSTM Architecture
Three Use Cases
More Expressive Architectures Make a Big Difference
Fixed Features Allow for Rapid Development
Our partners
Babel resources
Taught by
MITCBMM
