chatterbox-turbo-coreai-fp16

Chatterbox Turbo for Core AI: FP16 weights

English Chatterbox Turbo for OpenReader, with FP16 T3 weights and attention cache, and FP32 T3 intermediate math. S3Gen and the mandatory Perth watermark are the original FP32 exports. Other OpenReader TTS models are unchanged. This is a smaller download, not a claim of all-half arithmetic or realtime speech.

Source is resemble-ai/chatterbox at 5de7a54aa4e5e2baadb0182dde554908b48b85c2; weights, tokenizer and official default voice conditioning are ResembleAI/chatterbox-turbo at 749d1c1a46eb10492095d68fbcf55691ccf137cd. The MIT notice is in LICENSE.txt. There are no personal voice recordings.

T3's graph takes shared half-weight tensors as inputs. This avoids expanding constant casts during device specialization, while keeping FP32 normalization, projections, attention, residuals and output logits. runtime.json describes all inputs and verifies each weight tensor's shape and SHA-256. The native OpenReader runtime must support t3_weights; older builds cannot use this pack. The T3 program and its weight files together occupy about 751 MB versus 1,503 MB for the original T3 asset. The full pack is about 1.33 GB versus 2.08 GB. Working K/V state is 96 MiB versus 192 MiB; the reusable voice prefix is 33 MiB versus 66 MiB. These storage figures are not measured runtime RAM reductions.

The decoder uses the original masked 128/256/512 mel-frame vocoders, native byte BPE and sampling, two Euler steps, pitch/harmonic processing, STFT/ISTFT, fades and the learned Perth watermark. T3 has 1,024 cache positions and a 256-speech-token window including the three official silence tokens. Narration requires bounded passages. No Misaki/espeak runtime is involved.

Validation and device scope

The independently replayed PyTorch graph is byte-identical to the constant half-weight wider-math reference. On macOS Core AI CPU, strict export parity passed; two public texts also matched the original FP32 speech token IDs and waveform gates. Max PCM errors were below 0.000112 and repeated outputs matched exactly. This is numerical evidence, not a listening study.

On an iPhone 15 Pro (A17 Pro), Release OpenReader matched every converted T3 logit/token gate. Final PCM was byte-for-byte identical to the independent original FP32 phone pipeline for one fixed public fixture. The unchanged ANE-preferred S3/Perth path differs from the official CPU waveform under both precisions; that device comparison is separately identified in validation.json.

T3 remains CPU-only: ANE-preferred loading terminated the process on this phone. S3 and Perth keep the existing ANE preference. A short playback check completed with 6.31 s preparation and 15.20 s to first audio after preparation. Four-segment narration completed with 6.09 s preparation, 25.47 s to first audio, 24.88 s of audio and 111.06 s total. Sampled peak footprint was 2.18 GB on that passage, higher than the earlier FP32 passage; RAM did not halve. Turbo is still slower than realtime. Specialization caches were retained; these are not cold-start measurements from a wiped container.

files.json provides the exact payload sizes and SHA-256 hashes; validation.json records the measurements, rejected candidates and limitations. Reproduction is in OpenReader's tools/chatterbox and tools/tts_probes. The separate export environment pins Torch 2.9.1, coreai-core 1.0.0b3, coreai-torch 0.4.3 and coreai-opt 0.3.0; the unmodified official reference retains its Torch 2.6 environment.