Show HN: Drama 3 (35B TTS) speaking Turkish at 2.7x real time on one 16GB GPU (github.com)

A new release of Drama 3, a text-to-speech (TTS) model, has demonstrated impressive performance by speaking Turkish at speeds of 2.7 times real-time for short sentences and up to 3.3 times for longer paragraphs, all while running on a single 16GB GPU. This advancement retains the emotional nuances of speech thanks to a distilled architecture that utilizes 32 expert speakers and sophisticated memory management, where large model data is streamed directly from NVMe storage, preventing it from overloading RAM. This development is significant for the AI/ML community as it optimizes performance and resource utilization in TTS systems, opening doors for more efficient applications in real-time voice generation and multilingual support. Key technical innovations include improvements in processing time through compiled CUDA graphs, reduced load times for expert models, and enhanced emotional control via activation steering. The ability to distill a robust model that runs efficiently on limited hardware signifies a step towards making advanced AI speech technologies more accessible for developers and researchers, paving the way for broader applications in various linguistic contexts.