Evaluating the Impact of AI-Generated Data on Training Keyword Spotting Models

Keyword Spotting (KWS) systems have become common in everyday applications, used in virtual assistants (i.e. Amazon’s Alexa) and voice-controlled devices. These systems are typically part of complex architectures designed to simplify daily tasks and they work by continuously monitoring audio input for specific "wake words", such as "Hey Siri" or "Alexa", triggering an action or response when those words are detected through recognition models. A key challenge in designing KWS systems is the data collection process, which is often resource-consuming, especially when employing deep learning models since they require many high-quality recordings for effective training. The thesis examined the impact of blending well-known KWS datasets with synthetic samples generated by modern Text-To-Speech (TTS) systems. The objective was to determine whether integrating synthetic data could reduce the resources required for dataset construction while maintaining model performance. Several hybrid datasets were built, combining original KWS datasets with synthetic speech, and were then used to train state-of-the-art KWS models. Finally, the results were compared to models trained solely on original data. The experiments revealed a performance drop when synthetic data was introduced, with the decrease becoming more evident as the number of synthetic samples increased. This result highlights the need for particular care when using synthetic data to ensure the quality of KWS models is not compromised.