AuK Ships 1.5B Open Speech Model Unifying Generation, Editing

AuK, a 1.5-billion-parameter open-source model, unifies speech generation and editing behind one natural-language interface, and its authors report a 2.65 average word-error rate on Seed-TTS-Eval against 3.07 for Qwen3-TTS. The technical report describes a system covering five task families: speech generation, content editing, enhancement and separation, paralinguistic editing, and acoustic editing.

The training scale is heavy. The paper reports "approximately 3.03 billion instruction–audio instances and 1.95 million hours of effective supervision," with a transformer backbone of 10 dual-stream MMDiT blocks followed by 20 single-stream DiT blocks.

The editing numbers are where AuK separates from prior open models. On the Ming-Freeform-Audio-Edit benchmark's harder Chinese subset, it cuts word-error rate from 10.46 (Ming-UniAudio) to 3.09. On SpeechEditBench, content-editing accuracy climbs from 76.46 to 91.83.

Some behavior extends past the training tasks. "Our de-accenting supervision covers Chinese dialects and regional accents only, yet the model can reduce accents in English speech, including English spoken with Indian or Japanese accents, while largely preserving speaker identity," the authors write, calling out emergent whisper-style generation as another example.

A distilled variant, AuK-Flash, "performs 4-step inference without classifier-free guidance and achieves a 4.5× wall-clock speedup over the full model under matched conditions." Weights and code are released alongside the paper.

The release lands during a busy stretch for open voice models we've been tracking: Tencent's Gander shipped the same day, and VibeVoice-ASR posted 1.5B and 7B open weights days earlier.