Breaking

Tencent Releases AuK, an Open-Source Speech Model for Instruction-Based Audio Generation and Editing

September 15, 2026 at 11:16 EDT

Foundation Models
Media Generation
Open Source

Tencent Hunyuan has released AuK, a 1.5B-parameter foundation model that uses natural-language instructions to generate and edit speech, including changing spoken words, voice characteristics and emotional delivery. The project includes public code and model weights under an MIT license, with AuK and the faster AuK-Flash variant available through Hugging Face and ModelScope.

September 2026 · Tencent Hunyuan

One model for speech edits

AuK uses natural-language instructions to generate and edit speech, covering voice, wording, emotion, restoration and separation in a single open system.

1.5B

foundation model parameters

3.03B

instruction-audio instances

1.95M

hours of effective supervision

4.5x

AuK-Flash wall-clock speedup

AuK-Flash compresses inference steps

Full model

baseline

AuK-Flash

4.5x faster

The distilled variant runs 4-step inference without classifier-free guidance under matched test conditions.

One instruction field, many audio jobs

Prompt→GenerateEdit wordsChange emotionEnhanceSeparate

Why builders care

Open weights, an MIT license, demos, CLI tools, ComfyUI integration and fine-tuning scripts make the release easier to test and adapt.

What remains sensitive

Generated edits complicate provenance, voice authenticity and evidence workflows, while high-end inference still carries notable memory demands.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.