Advances in Signal Processing and Intelligent Recognition Systems by Unknown

Author:Unknown
Language: eng
Format: epub
ISBN: 9789811548284
Publisher: Springer Singapore

1 Introduction

Over 53 million people in the U.S. already own at least one smart speaker and it has increased 78% year-over-year as compared to 2017 [1]. Nowadays, with an enormous growth of gadgets, speech processing has become an important feature to enhance the human machine interaction. Speech processing is also used abundantly in chatbots for providing answers to queries asked by the customers [2]. This requires a large vocabulary speech recognition along with a robust language model. However, in several use cases, we need to recognize only words belonging to a certain limited vocabulary. For instance, voice enabled switch boards, smart home appliances, voice controlled car audio-systems, handling mechanical machines with speech command interface etc. Here, full-blown speech recognition system becomes an expensive affair requiring a lot of computing and communication bandwidth. Hence for these use cases, Speech Command Recognition (SCR) proves to be an efficient and reliable solution [2–4]. An ideal SCR system must be able to detect limited speech commands, with small footprint, high accuracy and minimal delay. It should also support multiple languages and dialects, should be noise-robust and speaker-independent.

In this paper, we propose a CNN based Streamed Speech Command Recognition (SSCR) system which works on single command as well as on stream of commands. Further, we enhance the dataset through speech augmentation techniques, as discussed in [5], by adding background noise, speed variation and time shifting. To demonstrate the utility of SSCR, we have developed end-to-end Voice Controlled Media Player (VCMP) controlled by 7 pre-defined commands. Rigorous quality assurance (QA) test was further carried out to verify its performance. We present comparative analysis of the two models (A) CNN model trained on Noise-free dataset (Noise-Unaware model) (B) CNN model trained on Noisy dataset (Noise-Aware model). Our key contributions are summarized below: 1.Convolutional Neural Network (CNN) model trained on Tensorflow dataset [6] for 10 commands like “yes”, “no”, “up”, “down” etc.Increased robustness of model by using data augmentation techniques.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.