AI4Bharat Public
Seminars
Seminars
Talks
Title
Date Time
Speaker
Talk Abstract
Speaker Bio
Slides
Meeting link
1
RNN-T ASR Systems and Enabling Contextualization For RNN-T ASR Systems
Mar 19, 9:00 AM
In this talk, Mahaveer will first discuss the general theory behind building Recurrent Neural Network Transducer (RNN-T) Automatic Speech Recognition (ASR) Systems. The RNN-T based End2End ASR systems has become a popular choice in the industry to build compact, accurate ASR systems that can be deployed on-device. Next, Mahaveer will discuss methods to enable contextualization for RNN-T ASR Systems. Contextualization allows us to use utterance specific context for ASR systems.
Open
2
RNN-T ASR Systems and Enabling Contextualization For RNN-T ASR Systems
Mar 26, 9:00 AM
In this talk, will discuss beam-search decoding and methods to enable contextualization for RNN-T ASR Systems
Open
3
No Language Left Behind: Unlocking Text Data for Under-Resourced Languages
Apr 7, 6:00 PM
Shruti Rijhwani
NLP systems are limited by the availability of text data, and because machine-readable text exists only in a few hundred languages, most of the world’s languages are under-represented in modern language technologies.
Text data exists in many more languages! However, it is locked away in printed books and handwritten documents, and training a high-performance optical character recognition (OCR) system to extract the text is challenging for most under-resourced languages.
In this talk, I will describe two methods for improving text recognition in low-resource settings using automatic OCR post-correction. The first is a multi-source encoder-decoder model with structural biases to efficiently learn from limited data. The second is a semi-supervised learning technique that uses raw unlabeled images to improve performance without additional manual annotation. The method combines self-training with automatically derived lexica through the use of weighted finite-state automata (WFSA) to improve post-correction. I will present empirical evaluation on multiple under-resourced languages to illustrate the effectiveness of the proposed approaches as well as future applications in using the extracted texts to expand multilingual NLP models to many more languages.
Open
4
Open
5
Open
6
Open
7
Open
There are no rows in this table
Want to print your doc?
This is not the way.
This is not the way.
Try clicking the ⋯ next to your doc name or using a keyboard shortcut (
CtrlP
) instead.