AuK

"An Open-Source Foundational Model for Speech Generation and Editing." Tencent Hunyuan's 1.5B AuK unifies zero-shot and instructed TTS, content and acoustic editing, paralinguistic editing, enhancement, and separation behind one interface of natural-language instructions plus audio context. It was trained on about 3.03 billion instruction-audio instances, 1.95 million hours of effective supervision across five task families: a multimodal LLM conditions the generator, a joint speech, audio, and music VAE provides the latent space, and a hybrid MMDiT-to-DiT rectified-flow transformer generates, followed by preference optimization and RL post-training. AuK-Flash is a four-step distilled variant about 4.5 times faster. Open-sourced September 9, 2026 under MIT (439 GitHub stars in its first days); the report posted September 8. The generation stack is trained from scratch; the conditioning encoder is a borrowed multimodal LLM.