Show HN: Homi, open decision model for audio, near Flash on MMSU and 50x cheaper (github.com)

A new open decision model called Homi has been introduced, enabling advanced analysis of audio clips by responding to typed questions. Unlike existing models like Jev, Clef, or pplx-decider that solely handle text input, Homi processes audio directly and provides probabilities for various emotional and phonological questions, such as determining speaker emotions. Homi outperformed Gemini Flash in VocalGrad metrics and is competitively close on the MMSU benchmark while being significantly cheaper—50 times less costly in operational expenses compared to Gemini. This development is significant for the AI/ML community as it opens new pathways for audio data evaluation at scale, tackling aspects of audio that text-based models cannot decipher, such as emotional tone or subtle phonetic cues. With capabilities to run three types of questions—yes/no, multiple-choice, and ordered scoring—Homi leverages state-of-the-art training techniques and data sourcing to generate precise audio assessments. Its cost efficiency and enhanced accuracy can lead to broader applications in fields like sentiment analysis, voice recognition, and AI-driven audio evaluation tools. The model's architecture and training process have been documented extensively, allowing for further innovation and development within the community.