OffNote AI Talks: Building Real-world ASR Systems | Speech Recognition | Speech-to-text
Abstract: This tutorial is an introduction to Automated Speech Recognition (ASR) for deep learning practitioners and enthusiasts. We’ll start with the basics of speech signal processing and build on it to cover the components of state-of-the-art ASR systems. This includes deep dive into some of the interesting techniques used in speech processing: CTC loss, 1-D time-channel separable convolutions, language model-based rescoring and beam search. Next, we’ll look into the challenges of building and deploying ASR systems in the wild, and some solutions to address these. The challenges include handling of low resource languages, data collection and cleaning, handling streaming audio and cost-effective deployment. The solutions include transfer learning, audio augmentations and proportional sampling-based training schemes. Finally, we’ll discuss a few interesting directions for further R&D in ASR.
Speaker: Harisankar Haridas (Airtel X Labs)
Host: Nishant Sinha (OffNote Labs)
Presented on: 23rd November 2020
Materials/References:
Slides, Notebook, Sample Audio: https://voicetech.substack.com/p/comi...
Concepts covered:
Deep learning-based ASR
Basics of audio signal processing
CNNs for deep learning (Quartznet)
CTC loss, 1-D time-channel separable convolutions
Language models for ASR
Beam search
Low-resource languages
Code-mixing
Speech data collection: methods and tools
Speech data cleaning
ASR on streaming audio
Pointers to open source tools
Feedback
Please give your feedback on the tutorial in this form: https://docs.google.com/forms/d/e/1FA...
===
The OffNote Labs AI Talk Series brings you industry experts, researchers and practitioners, passionate to share their learnings and experience -- on innovating and building cutting-edge AI technology / systems which touch and influence the lives of billions of people on this planet.
Our talks are informal, a blend of traditional presentation / podcast, and the audience very technically engaged.
We take delight in unraveling the experience of research and innovation, and celebrate innovators who 'follow the problem', and make complex technology work in the wild.
==
Follow OffNote Labs on LinkedIn. / offnote
Watch previous talks and Subscribe on Youtube: https://bit.ly/31VTMHH
Sign up to our newsletter: http://offnote.substack.com
Read our research articles: / offnote
Web: https://offnote.co
===
00:00 Introduction
04:48 Overview - Building Real-world ASR Systems
05:55 Relevant Background
07:02 Talk Outline
08:53 Demo
10:22 ASR system
24:48 Demo: One Char per Time Step
31:06 Quartznet Acoustic Model (2019)
33:20 1D Time-Channel Separable Convolutions
38:08 CTC Loss
47:02 Language Model
47:46 Need for Language Model
52:57 Beam Search
58:31 Part 1- Q&A
1:1151 How to Address Practical Challenges
1:14:26 Audio-text Paired Data: English
1:17:02 Audio-text Paired Data: Hindi
1:18:31 Speech Data Characteristics
1:22:35 Hindi Data Collection
1:28:53 Text Corpus for Data Collection
1:31:51 Clean the Text Data
1:38:31 Handling Code-mixing ( Hindi- English )
1:44:04 Augmentations
1:46:17 Specaugment (2019)
1:51:51 Audio-text Paired Data: Hindi - Challenge
1:55:55 Forced Alignment
2:09:37 Transfer Learning
2:12:13 Transfer Learning: Sample Results
2:14:04 Accuracy Benchmarking
2:15:55 Handling Different Speech Segments
2:20:00 Further Challenges
2:21:28 Handling Short/Streaming Audio
2:23:20 Interesting Directions for further R&D
2:25:33 Other Resources
2:28:08 Conclusion
#asr #speechrecognition #speechtotext #deeplearning #offnote #research #indic #languagemodel #beamsearch #paireddata #specaugment #forcedalignment #transferlearning
#speechsegment #quartznet #acoustic #airtelXlabs