Smart Turn v3.1, an upgraded version of the Smart Turn model, has been released with significantly improved turn detection accuracy, utilizing a larger dataset of human audio samples contributed by partners such as Liva AI, Midcentury, and MundoAI. This version maintains the same model architecture as v3.0 and can be used with existing inference code by replacing the ONNX file, ensuring it serves as a drop-in replacement. The improvements are attributed to the inclusion of authentic human audio data, which enhances the model's ability to capture natural speech variability, marking a notable advancement from the synthetic data previously relied upon. Smart Turn v3.1 is available in two variants, an 8MB CPU model with fast inference times and a 32MB GPU model that offers slightly better accuracy. Released as open-source, this update supports 23 languages, with notable accuracy gains in English and Spanish, and is integrated into the Pipecat platform to facilitate easy adoption. Future development will continue to focus on refining accuracy across all languages while maintaining open access to models and datasets for community collaboration.