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

Category: ai-ml

Tags: text-to-speech, turkish, model-optimization, gpu-inference, moe-distillation

Score: 7.3/10 (Innovation: 7, Technical: 8, Documentation: 8, Utility: 6)

This project adapts Fish Audio Drama 3, a 35B parameter text-to-speech model, to speak Turkish on a single 16 GB GPU. It achieves 2.7x real-time performance through expert distillation, hot-expert caching, disk streaming, and CUDA graph optimizations, with detailed benchmarks and an emotion intensity control via activation steering.

Target audience: ML engineers, TTS researchers, and developers working on low-resource language speech synthesis

Repository: https://github.com/kasparovabi/drama-3-tr · Python · Apache-2.0

View on Hacker News