| Chatterbox TTS — CoreML Port |
| ============================= |
|
|
| This repository contains a CoreML conversion of the Chatterbox TTS model |
| originally released by Resemble AI. |
|
|
| Upstream |
| |
| Project: Chatterbox |
| Author: Resemble AI |
| Source: https://github.com/resemble-ai/chatterbox |
| License: MIT (see LICENSE) |
| Copyright: Copyright (c) 2024 Resemble AI |
|
|
| This port is distributed under the same MIT License. The full upstream |
| copyright notice and license terms are reproduced verbatim in LICENSE, as |
| required by the MIT License. |
|
|
| Modifications |
| |
| The following changes were made to the original model artifacts to produce |
| the files in this repository. The modifications cover only the |
| serialization/inference graph; the trained weights are preserved unchanged. |
|
|
| * Converted PyTorch / ONNX checkpoints to CoreML `.mlpackage` format using |
| Apple `coremltools`. |
|
|
| * Replaced `complex64` rotary position embeddings (RoPE) inside the s3 |
| tokenizer with split real `(cos, sin)` tensors. CoreML does not support |
| `complex64`; this rewrite is mathematically equivalent. |
|
|
| * Traced the flow encoder at fixed shapes (400 prompt tokens, 1024 mel |
| frames). The host application is expected to chunk longer inputs to |
| these limits. |
|
|
| * Pre-computed and serialized the conditioning tensors for the bundled |
| default voice (the five `default_*.bin` files) so callers can run |
| inference without re-running the voice encoder pipeline on every cold |
| start. |
|
|
| The conversion script (`convert_chatterbox_to_coreml.py`) lives in the host |
| application's source repository. No model weights were retrained or |
| fine-tuned; only the serialization format was changed. |
|
|
| Trademarks |
| ---------- |
| "Resemble AI" and "Chatterbox" are referenced here for attribution only and |
| remain the property of their respective owners. Their inclusion does not |
| imply endorsement of this port. |
| |
