Tencent open-sources "Audio Version of Nano Banana": covers text modification and voice replacement, voice editing tools surge in rankings.

Beating AI Express: Tencent Hunyuan has open-sourced AuK, a 1.5-billion-parameter speech generation and editing model, which the official dubs the "audio version of Nano Banana". The model integrates functions including voice cloning, word editing, emotion switching, accent removal, speed and pitch adjustment, noise reduction, and multi-person/music separation, all controllable via natural language. AuK can directly modify existing recordings: it only adjusts incorrect words without re-recording the entire segment, alters emotion/timbre/accent while keeping content unchanged, edits lyrics, removes laughter/coughs, and converts normal speech to whispers. Notably, its voice cloning does not require a reference transcript—only a sample audio and new text are needed. In official tests, AuK leads in multiple speech generation and general speech editing benchmarks, scoring 49.73 on SpeechEditBench, while second-place Ming-UniAudio scores 28.70. The fast variant AuK-Flash, after distillation, needs just 4 inference steps and runs ~4.5x faster. However, the paper notes AuK cannot yet reliably understand all arbitrary natural language inputs, still relying on a Prompt Enhancer to identify tasks, organize parameters and rewrite instructions. Code and model weights are now public under the MIT license.