Kokoro TTS ported to Apple Core AI: 54 voices running fully on-device with zero API cost
amos_gyamfi · x · 2026-09-14
A developer converted the open-source 82M-param Kokoro-82M TTS model (StyleTTS2 + iSTFTNet, Apache-2.0) into Apple's new on-device runtime Core AI (.aimodel format), offering 54 male and female voices that run fully locally on iOS/macOS — no API calls or token costs.
Key points:
- Non-autoregressive architecture: phonemes + a voice/style vector generate the waveform in one pass, 24 kHz, English-first
- Runs on GPU or Neural Engine via Core AI (successor to Core ML in iOS 27 / macOS 27); the same benchmark protocol shows Qwen3-8B 4-bit at 94 tok/s on an M4 Max GPU
- One-line usage: CoreAI.speak(text, options: .model("kokoro-82m")), with on-first-use download; a full Speak example app and CLI are available in the coreai-kit repo
- Complete Swift snippets using CoreAIKit make it copy-paste runnable
More from Infra
- 15 self-funded GPU jobs show cost estimators overshoot by a third — Worldly_North_7213 · 2026-09-14
- Speculative decoding on or off: the 35B MoE offloading question on 8GB VRAM — Infinite-Local5435 · 2026-09-14
- Running an LLM agent on a 512MB board with decoupled memory and live cross-machine migration — D777Castle · 2026-09-14
- Across 3,171 sessions and 30B tokens, only 0.3% was model output — a local tool that searches your agent logs — Rare_Guide_9830 · 2026-09-14
- Maia 200 hits ~12 TFLOP/s FP4 in 1mm²: density should be a first-class goal — thoefler · 2026-09-14
- Dev's 24/7 self-hosted AI stack: OpenWebUI, pidot, Tailscale, GLM and DeepSeek — andfanilo · 2026-09-14