SpeechTera Ltda: development of computational resources for speech technologies

• Grant number: 16/08355-0
• Support Opportunities: Research Grants - Innovative Research in Small Business - PIPE
• Start date: April 01, 2017
• End date: August 31, 2019
• Field of knowledge: Interdisciplinary Subjects
• Principal Investigator: Vanessa Marquiafável Serrani
• Grantee: Vanessa Marquiafável Serrani
• Company: SpeechTera Desenvolvimento de Programas para Computadores Ltda
• City: Araras
• Associated research grant: 14/21750-0 - Speech Tera Ltda: development of computational resources for speech technologies, AP.PIPE
• Associated scholarship(s): 18/06601-0 - SpeechTera Ltda: development of computational resources for speech technologies, BP.TT; 18/06738-5 - SpeechTera Ltda: development of computational resources for speech technologies, BP.TT; 18/02617-9 - SpeechTera Ltda: development of computational resources for speech technologies, BP.TT; 17/07371-5 - SpeechTera Ltda: development of computational resources for speech technologies, BP.PIPE

Abstract

This project aims to create computational resources for the development of Speech Technologies, focused on Brazilian Portuguese. With the development of robust algorithms to treat speech databases, applications involving recognition or speech synthesis, respectively, ASR (Automatic Speech Recognition) and TTS (Text-to-Speech), have gained more space in our everyday life and become increasingly accurate. However, although Brazilian Portuguese is the sixth most spoken language in the world, resources availible for the development of speech processing technologies for that for that language are scarce: there are few databases, grapheme-phoneme converters and acoustic or pronunciation models on the market. This project seeks to act precisely in that gap. Our purpose is to develop computational resources in order to encourage the development of speech technologies for Brazilian Portuguese, both for the industry and academia. Its proposes are the development of four types of products: i) speech corpora ii) acoustic models, iii) models of pronunciation and iv) grapheme-phoneme converters. For speech corpora, we propose methods of collecting and annotating data based on crawling and crowd-sourcing, that will enable the development of speech resources at the most competitive and affordable prices that currently in existence on the market. State of the art techniques will be employed in the preparation of the acoustic models, like Deep Neural Networks; and grapheme-phoneme converters as hybrid models based on manual rules and machine learning techniques (SVM, CART, MARS). The proposed business model focuses on a business-to-business approach (B2B), focused on Information Processing; Speech and Natural Language Processing technology companies, especially with the startups niche in mind. (AU)