Data-Centric Support for Modeling Spoken Queries on Virtual Assistants
DSc. Vítor Silva Sousa (Apple Inc) 

Data-Centric Support for Modeling Spoken Queries on Virtual Assistants   DSc. Vítor Silva Sousa (Apple Inc.)    Vitor Silva Sousa será apresentado pela Prof. Marta Lima de Queirós Mattoso   Dia 06/11 (quarta-feira), 18 horas, Transmissão ao vivo no Canal do PESC no Youtube

Abstract:

Virtual Assistants (VAs) are becoming Information Retrieval (IR) platforms, where users utter queries as voice commands possibly preceded by a wake word such as “Hey VA". Despite the complexity of Automatic Speech Recognition (ASR) systems on VAs, they still face challenges associated with accurate spoken voice query transcription. Language Models (LM) trained on text assist in handling ambiguous and difficult-to-understand queries. Nevertheless, LMs require external data sources, e.g., Knowledge Graphs (KG), query templates, and popularity signals that can help boost trending/popular entities. Regarding operations, linguists routinely need to investigate upstream data manually to understand the performance of their trained LM, whose solutions are not offered as off-the-shelf software. Therefore, data-centric support is paramount to providing analytical capabilities, debugging potential data incidents, deciding on LM deployment, and monitoring performance over time. In this talk, Vítor will present a data-centric approach for ASR systems. 

Short Bio: 

Vítor Silva is AI/ML Senior Speech Data Engineer on the Siri Understanding team at Apple Inc. He received his Ph.D. in Systems Engineering and Computing at PESC/COPPE/UFRJ. Vítor has experience in machine learning and large-scale data engineering in industry. He worked as a Senior Engineering Technologist at Dell EMC and a Research Engineer as part of the Computational Social Science team at Snap Inc. He has published 68 scientific papers, and he holds 10 US patents on the following domains: machine learning, large-scale data engineering, and provenance.