Audio8 TTS: open 0.6B zero-shot voice cloning, runs on a laptop CPU
Apache-2.0 weights, 11 languages, ~1GB via ONNX INT4 — and GitHub Copilot retires six models on Sept 1.
Copy markdown
Clone any voice from a few seconds of audio
Audio8 TTS Preview is an open (Apache 2.0) 601M-parameter model that does zero-shot voice cloning: hand it a short reference clip and it speaks in that voice across 11 languages, including English, Chinese, Japanese, Korean, and German. Output is 44.1kHz.
Small enough for a MacBook — or a bare CPU
It ran on a 16GB M2 MacBook Pro with no swapping, and peaked around 2.9GB of memory on an RTX 3090 Ti for a two-item batch. An ONNX INT4 build fits in roughly 1GB, so you can run it CPU-only with no GPU at all.
SOTA-class accuracy at a fraction of the size
On the Seed-TTS benchmark it posts a 1.506 English word-error rate, competitive with far larger systems despite being the smallest model in its comparison group, plus a 0.950 Chinese character-error rate. It uses a DualAR architecture inspired by Fish Audio S2 Pro.
Build this weekend: a private narration or dubbing tool
Batch inference runs from a JSON manifest and an SFT training pipeline ships in the repo, so you can fine-tune on your own voice and generate narration or dubbing entirely offline. Caveat: it's a preview, so languages are intentionally limited and your reference transcript must match the reference audio.
Elsewhere: Copilot drops six models on Sept 1
GitHub Copilot is retiring Claude Opus 4.5 and 4.6, Sonnet 4.5 and 4.6, Gemini 3.1 Pro, and Raptor Mini on Sept 1 — move to Opus 4.7/4.8/5, Sonnet 5, Gemini 3.6 Flash, or MAI-Code-1-Flash. Annual individual subscribers keep Sonnet 4.6.