The Key Role of Data in Modern AI-Driven Systems: Spotting Voice Keywords and Beyond

Datasets are essential to AI models.  They provide the truth by which we train AI models and the tests by which we measure AI success.  While researchers tend to reuse well-known datasets, engineers building real-world systems must create datasets that represent all scenarios in which the AI model is expected to operate. This is often an iterative process that requires application-specific resources, tools, and expertise.

In this session, we will explore a well-known practical example: waking up voice-enabled devices using trigger phrases like "Hey Siri" or "OK Google." We will cover a number of data-specific best practices focused on data labeling and annotation, data ingestion, data synthesis and augmentation, feature extraction, and domain transformations. This practical example provides general considerations that can be applied to a wide range of applications.