Fourteen MB Hinglish STT pWER 29, stary words 2.3/100 (hitl.systems)

The release of a compact 14.4 MB speech-to-text model, lipi-hn, specifically designed for Hinglish—an Indian code-switching language blending Hindi and English—marks a significant advancement for the AI/ML community. With a promising performance, achieving a 29.0 pWER (Phone Error Rate), it surpasses larger counterparts, such as larger Whisper models, which recorded 39.0 and 68.8 pWER. The model is optimized for offline use on CPU-only devices and outputs in Roman script, consistent with how users naturally type, making it especially suitable for apps targeting Indian users. This development is notable not only for its small size and efficient decoding options but also for its accessibility; it is free for non-commercial use under a CC BY-NC-SA 4.0 license. The model provides a robust solution for developers looking to integrate Hindi-English speech functionalities into applications, allowing them to deliver a seamless user experience without the need for a continuous internet connection. However, it is notably intended for non-commercial projects, highlighting its potential to support open-source, research, and accessibility initiatives within the AI space.