Reproduce “Training Keyword Spotters with Limited and Synthesized Speech Data” on OpenResearch

About this paper

James Lin, Kevin Kilgour, Dominik Roblek, Matthew Sharifi

Google Research developed a method to efficiently train small, on-device keyword spotters by combining a large-scale pre-trained speech embedding model with high-quality synthesized speech, significantly reducing the reliance on extensive real-world training data. This approach enabled the creation of effective keyword spotters with accuracy comparable to models trained on substantially more real examples.

Key insights

• A general-purpose speech embedding model, pre-trained on a massive dataset, can significantly reduce the amount of task-specific real data required to train new keyword spotters.
• High-quality synthesized speech effectively substitutes for real audio when combined with pre-trained embeddings, making keyword spotter development feasible in data-scarce scenarios.
• Adding even a small amount of real speech data to synthetic datasets substantially improves accuracy, bridging the gap between purely synthetic and purely real-data training.

Summary by alphaXiv.