On-Device Neural End-to-End Speech Recognition and Synthesis Algorithms - A Review

Explore a comprehensive review of on-device fully neural end-to-end speech recognition and synthesis algorithms in this 49-minute plenary talk from tinyML Asia 2021. Delve into the evolution from conventional speech recognition systems to modern fully neural network-based approaches. Examine various end-to-end automatic speech recognition and speech synthesis algorithms, including CTC, RNN-T, AED, MoChA, and transformer-based systems. Learn about the challenges of implementing traditional systems on devices and how neural network-based solutions offer smaller memory footprints. Discover advancements in Text-to-Speech (TTS) technology, from parametric and concatenative approaches to neural speech synthesis methods like Tacotron and Wavenet. Gain insights into model compression techniques, custom recursion, and hybrid approaches for improving speech recognition latency. Investigate the potential of NPUs and applications for low-end devices, and explore future possibilities in the field of speech technology.

Introduction
Contents
Conventional techniques
Commercialization
Special Recognition Systems
TC Loss
AttentionBased Approach
AttentionBased Problem
Monotonic Trunkwise Attention
Model Compression
Custom Recursion
Speech Request Latency
Hybrid Approach
Summary
NPU
Low end devices
Future applications