Developer runs 35B Drama 3 voice model on a 16 GB RTX 5070 Ti

Developer Ahmet Kazankaya adapted Fish Audio's 35B-parameter Drama 3 text-to-speech model to Turkish on a 16 GB graphics card. He split reading from speech generation, trained a compact 32-of-256-expert speaker and streams expert weights from NVMe, reaching 2.68% WER on the Freya-TR-Eval benchmark.

• The 35B model needs 41.5 GB for weights but runs on 16 GB of VRAM
• Speaker trained on 32 of 256 experts per layer in 85 minutes on 3,800 utterances
• 2.68% WER and 0.70% CER across 495 sentences on Freya-TR-Eval
• Disk-streamed experts slow short sentences to 0.86x real time