Tencent Releases Ultra-Fast Voice Editing AI AuK-Flash

Tencent has released AuK-Flash, a distilled version of AuK, a 1.5B parameter foundation model for voice generation and editing. This model can perform various voice tasks through natural language instructions, and its inference speed has been improved to enable high-quality voice generation with only 4-step inference through distillation of the original model.

Key Features and Performance

• Ultra-fast inference: AuK-Flash is a distilled version of the original model, enabling voice generation with just 4-step inference, significantly improving inference speed.
• Unified interface: All tasks, including zero-shot TTS, instruction-based TTS, content editing, acoustic editing, emotion editing, voice enhancement, and source separation, can be handled with a single natural language instruction.

Supported Task Scope

AuK-Flash supports a wide range of voice manipulation features:

• Voice generation: Zero-shot TTS using reference audio, Instruct TTS generating from voice descriptions alone
• Content and acoustic editing: Text replacement/insertion/deletion, pitch/speed/volume adjustment, lyrics editing
• Paralinguistic editing: Emotion change, timbre change, accent removal, addition/removal of non-verbal sounds (breathing, laughter, etc.), whisper conversion
• Enhancement and separation: Noise removal and reverb removal, speaker separation, vocal extraction from music, specific speaker extraction

Weights are available for download on Hugging Face and ModelScope, and CLI and Python example code can be found via GitHub and the project page.