Improved End-of-Turn Model Cuts Voice AI Interruptions 39%

The release of the transformer-based end-of-turn detection model version 0.4.1-intl marks a significant advancement in voice AI by enhancing accuracy and responsiveness across multiple languages. This update focuses on reducing false-positive interruptions and improving the handling of structured data, such as phone numbers and credit card details, by leveraging a large language model (LLM) backbone that combines semantic content and context. The model shows a 39.23% relative reduction in interruptions compared to its predecessor, with consistent improvements across languages like Chinese, Dutch, and Spanish. Enhanced training strategies, dataset composition, and preprocessing contribute to these achievements, while the adoption of a multilingual model replaces the legacy English model for broader applicability. The model's robustness is further enhanced by adapting to variations in speech-to-text outputs and integrating observability features for easier debugging. Future iterations aim to incorporate raw audio features to refine voice AI interactions, with the ultimate goal of creating more natural and human-like conversational experiences.