Tencent open-sources AuK: one 1.5B model generates and edits speech
It collapses TTS, denoising, speaker separation and prosody editing into one MIT model. Plus: GLM-5.3-Flash's promo just ended and Claude Code adds /diff.
Copy markdown
One model, your whole audio stack
AuK runs five job families from a single natural-language interface: zero-shot TTS, content editing (rewrite words or lyrics), acoustic edits (pitch, speed, volume), paralinguistic edits (emotion, timbre, accent, whisper), plus denoise and speaker separation. It's MIT-licensed, so commercial use needs no negotiation.
AuK-Flash: comparable quality in 4 steps
The distilled AuK-Flash matches base quality in 4 inference steps with a ~4.5x wall-clock speedup and no classifier-free guidance. Base and Flash weights are live on Hugging Face (tencent/AuK, tencent/AuK-Flash) and ModelScope, and run on a single GPU.
Build this weekend: a one-pass voice cleanup
Feed AuK a messy recording and get back a finished clip: strip room echo, split overlapping speakers, fix a misread word, then re-cast the delivery to a target emotion, all in one pass instead of chaining four separate models. Enough for a podcast cleaner, a dubbing tool, or an emotion-controllable narrator.
Heads-up: GLM-5.3-Flash's launch discount just expired
As of 16:00 UTC today, Z.ai's 50% promo ended: GLM-5.3-Flash now lists at $0.15 in / $0.50 out per 1M tokens ($0.03 cached). The natively-multimodal 320B/18B-active model carries a 1M-token context and ships open weights (MIT) at zai-org/GLM-5.3-Flash for local SGLang/vLLM serving.
Elsewhere: Claude Code 2.1.265 adds /diff and plugin folders
This week's Claude Code adds a fullscreen /diff to review uncommitted changes, --plugin-dir to load a whole folder of plugins at once, a 1 GB cap on tool results written to disk, and /skill-doctor to surface unused skills. A quick 2.1.266 hotfix restores standalone CLAUDE_CODE_USE_GATEWAY behavior.