Edit what a recording says, then change its mood, speed or accent

You give it a recording and a plain-language instruction, and it returns new audio. The instruction can be "replace 'Tuesday' with 'Thursday'", "make this sound cheerful", "slow it down a little", "remove the breaths", "take out the regional accent", or "keep only the second speaker". Every task goes through the same instruction box, so there is no separate tool for each job.

It also generates speech from scratch in two ways: clone a voice from a short reference clip and have it read your text, or describe the voice you want in words ("calm, low, middle-aged man") and get speech with no reference at all. On the cleanup side it denoises, removes room echo, pulls one voice out of a crowd, and separates vocals from a music mix. Singers get a lyric editor that rewrites the words while keeping the melody.

This is the Flash variant: the same model distilled to run in four passes instead of many, so a sentence renders in seconds. Tencent's own charts put the base model ahead of other open speech models on generation, editing, enhancement and separation, and the Flash version close behind. It speaks English and Chinese. The weights are MIT, so you can build a product on them.

Open Tencent's official demo Space, upload a short voice memo, and ask it to swap one word for another. Play the original and the edit back to back and listen for the seam; that single test tells you whether it holds up for your voice. Then try an instruction edit like "make it sound tired" on the same clip. To run it yourself, the GitHub repo ships a Gradio app, ComfyUI nodes, and a cookbook with the exact instruction template for each task; you download AuK-Flash plus the Qwen2.5-Omni-3B encoder it uses to read your instruction, and a single consumer GPU is enough for the 1.5B model.

The comparison charts are Tencent's own, and the model is five days old, so expect rough edges on voices and accents outside English and Chinese. Flash trades a little quality for speed; if an edit sounds off, the base AuK weights are the slower, cleaner option. Voice cloning from a short clip works on anyone's voice, so use it only on voices you have permission to use.

• Patch a mispronounced name or a wrong date in a podcast or voiceover without booking a re-take.
• Narrate a video in your own voice from a two-minute sample, then fix lines by editing text.
• Prototype a character voice from a written description before hiring a voice actor.
• Clean up an interview recorded in a noisy cafe and keep only the guest's track.
• Change the lyrics of a demo vocal while the melody stays where it was.

Try the demo

View model page

Read this issue on neuralexpedition.com →