Chatterbox Turbo for Core AI

FP32 Core AI conversion of Resemble AI's English Chatterbox Turbo, upstream weights revision 749d1c1a46eb10492095d68fbcf55691ccf137cd and source revision 5de7a54aa4e5e2baadb0182dde554908b48b85c2. Attribution and the MIT notice are in LICENSE.txt. This repository includes the official default conditioning, GPT-2 tokenizer assets and the mandatory Perth implicit watermark encoder. It contains no personal voice recordings.

The runtime has three shared-weight assets: T3 prefill/decode, S3 preparation/flow/pitch/harmonic merge/vocoder, and Perth. Total payload is approximately 2.08 GB. files.json records exact bytes and SHA-256 hashes. T3 has 1024 cache positions and 128 text-token capacity; the S3 window is 256 speech tokens including the three official silence tokens. Longer text must be chunked; reaching the speech limit without EOS is an error.

The vocoder uses 128/256/512 mel-frame buckets. Every intermediate convolution is masked at the actual boundary, preserving the unpadded official computation. A flexible-shape vocoder was rejected because repeated identical inputs produced unstable output. The other variable-length S3 functions retain actual shapes.

The native Swift implementation supplies byte BPE, token sampling, both Euler updates, harmonic construction, STFT/ISTFT, fades and Perth's 24/32 kHz host processing. The learned watermark is retained. These assets are not standalone Core ML packages or a hosted inference endpoint.

Measured on macOS 27 arm64 using Core AI CPU-only: three official reference passages exercise all vocoder buckets. All sampled token IDs match. Final watermarked output has matching lengths (29,760, 67,200 and 153,600 samples), maximum absolute error below 0.000036 and RMS error below 0.0000017. validation.json contains the measured results. The native implementation also compiles for iOS 27 in Release mode. Physical-device speed, memory, Neural Engine parity and background playback have not yet been measured.

Reproduction scripts and Swift sources are in OpenReader's tools/chatterbox and OpenReaderKit/Sources/Inference/ChatterboxTurbo. The reference environment uses official Torch 2.6; export uses Torch 2.9.1 and coreai-torch 0.4.1. Exported functions require the native state-management and audio pipeline; padding an ordinary unmasked vocoder does not reproduce the boundary computation.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for richardr1126/chatterbox-turbo-coreai

Finetuned
(13)
this model