Chatterbox Multilingual — CoreML

CoreML export of ResembleAI/chatterbox multilingual (23 languages, t3_mtl23ls_v2 + s3gen) for Apple platforms, converted by FluidInference (conversion toolkit: mobius PR #89).

Each model ships as both .mlpackage (source) and compiled .mlmodelc.

Models

File	Size (fp16)	Role	Compute
`T3-Prefill-T256-M1024-fp16`	977 MB	Llama-520M prefill over ≤256-token context (CFG batch 2), initializes 1024-slot KV cache	CPU+GPU
`T3-Decode-M1024-fp16`	977 MB	Single-step AR decode, KV cache via I/O tensors (38 ms/step)	CPU+GPU
`T3-Decode-M1024-fp16-stateful`	977 MB	Single-step AR decode, KV cache in `MLState` (16.8 ms/step; macOS 15+/iOS 18+)	CPU+GPU
`Flow-N500-fp16`	229 MB	S3Gen flow: 500-token bucket → 1000 mel frames, 10-step CFG Euler in-graph	CPU+GPU
`HiFT-T1000-fp16`	40 MB	HiFTNet vocoder: mel → 24 kHz waveform (0.09 s/call)	CPU+GPU
`tables/tables.safetensors`	34 MB	text/speech embedding + learned positional tables (host applies)
`tables/voice-default.safetensors`	0.1 MB	precomputed built-in voice conditioning (T3 cond embeds + S3Gen ref dict)
`tokenizer/grapheme_mtl_merged_expanded_v1.json`		23-language grapheme tokenizer

⚠️ Do not load the T3 packages with .cpuOnly — prediction hard-crashes (also independently reported by other Chatterbox CoreML ports). Use .cpuAndGPU or .all.

Samples

samples/ has CoreML end-to-end renders (e2e_*.wav) next to stock PyTorch renders (baseline_*.wav) for en/de/fr, all using the built-in voice.

To synthesize locally without the upstream checkpoint (Apple silicon):

git clone -b feat/chatterbox-mtl-coreml https://github.com/FluidInference/mobius
cd mobius/models/tts/chatterbox/coreml
uv sync
uv run python verify/e2e_coreml.py --lang en   # models auto-download from this repo

Runtime boundary

The graphs cover T3 prefill/decode (with the multilingual alignment-analyzer attention rows as outputs), the S3Gen flow, and the HiFT vocoder. The host runtime must provide:

text normalization + tokenization (tokenizer/)
embedding prep from tables.safetensors (text/speech + positional; the stock prefill context ends with two BOS embeds — replicate exactly)
CFG combine cond + w*(cond-uncond), repetition penalty, min-p/top-p, sampling, EOS handling
the AlignmentStreamAnalyzer heuristics, fed by the exported align_attn rows (reference port: verify/analyzer_port.py in the conversion toolkit)
SineGen randomness (phase_vec, noise inputs to HiFT) and CFM noise z
flow bucket padding/cropping; MLState seeding from prefill KV for the stateful decode

Voice cloning from a reference wav additionally needs the VoiceEncoder / S3TokenizerV2 / CAMPPlus encoders, which are not converted here; voices can be prepared offline in Python (export-tables.py --ref-wav) and shipped as voice-*.safetensors.

Parity (vs upstream PyTorch)

Check	Result
T3 wrappers vs stock (fp32)	logits 3.8e-05, alignment rows exact
T3 CoreML fp16	logits 2.6e-02 (range ±15), align 1.4e-03
Flow CoreML fp16	mel max 2.4e-02, mean 2.7e-03
HiFT CoreML fp16	wav max 1.7e-02, mean 2.6e-04
e2e ASR round-trip (en)	exact transcript, matches PyTorch baseline

License