ronbakker/chatterbox-nano-fp16-coreml

Chatterbox Nano Core ML

English speech synthesis and voice enrollment models for Apple Silicon, converted and optimized by Ron Bakker from original Resemble AI Chatterbox Nano source and checkpoints.

Models	Size
Narration, tokenizer and embedding tables	563.0 MB
Voice enrollment	182.7 MB
Optional zero-excitation vocoder	35.3 MB

Sizes are decimal MB. T3 uses grouped 8-bit weight palettes with FP16 computation and state. Acoustic models use FP16 weights. The default vocoder retains pitch prediction and harmonic excitation, with FP32 phase accumulation. Enrollment retains FP32 computation with compressed speech-tokenizer weights.

Requires iOS 18+ or macOS 15+ on Apple Silicon. Output is mono 24 kHz PCM. See INTEGRATION.md for model interfaces and conditioning. No voices or recordings are included.

Narration was tested on iPhone 16 Pro and M1 Max. Enrollment has component validation on Mac; validation of these enrollment exports on iPhone is pending.

MIT. Original model/source: Copyright (c) 2025 Resemble AI. Core ML conversion, optimization and packaging contributions: Copyright (c) 2026 Ron Bakker. This notice covers those contributions, not ownership of the original trained weights. Preserve LICENSE, NOTICE.md, and the included Apache-2.0 component license. Use reference recordings with permission.