rahulrachuri/pocket-tts-coreai

pocket-tts for Apple Core AI

Converted Core AI bundles for Kyutai's pocket-tts, running fully on-device on iPhone and Apple silicon Macs. Paired with the Swift host at RahulRachuri/pocket-tts-swift.

Source checkpoint: kyutai/pocket-tts-without-voice-cloning, revision e041936c75475d350b405bc870bcf7c22da4e9e6.

Performance

iPhone 17 Pro Max (iPhone18,2, A19 Pro), iOS 27.0, Release build, charging, thermal nominal. 148-word paragraph, one warmup and three timed runs, model load excluded.

config	RTF median	×realtime	peak RSS	load
fp16 gpu	0.1281	7.8×	169 MB	0.4 s
fp32 gpu	0.1646	6.1×	202 MB	0.7 s

For reference, on the same phone the Core ML route (FluidInference/pocket-tts-coreml through FluidAudio SDK 0.15.5 at its shipping defaults) measures RTF 0.399, or 2.51× realtime, on a length-matched passage.

Accuracy

Gated against the PyTorch fp32 oracle rather than judged by ear.

• End-to-end oracle prompt, device GPU: cos 1.000000, max|Δ| 4.3e−5, framing identical.
• Per-graph transfer under cpuOnly, device versus M4 Pro: bit-identical, max|Δ| = 0 across all eight dumps.
• ASR round trip: 0.00% WER on the oracle prompt, 1.38% on the 148-word paragraph, 2.45% on a 163-word, 47-second passage.

Contents

bundle	precision	notes
flowlm_float32_s512	fp32	flow-LM, multifunction (prefill + step), in-graph state
flowlm_float16_s512	fp16	same, half precision
flow_decoder_float32_lsd1	fp32	flow head
flow_decoder_float16_lsd1	fp16	same, half precision
mimi_decoder_float32_ring272_outer_q_gs	fp32	Mimi decoder, quantizer folded, in-graph ring state

These are JIT .aimodel bundles. For iOS, compile them AOT with --expect-frequent-reshapes omitted, since every shape here is static.

The model weights themselves are not mirrored here. The Swift host loads model.safetensors, tokenizer.model, and embeddings/<voice>.safetensors directly from Kyutai's checkpoint, which is ungated. Fetch it with huggingface_hub at the revision above.

Requirements

Exported with coreai-core 1.0.0b2. Bundles built with earlier versions are rejected by the Xcode 27 loader, and the export gives no warning when it happens.

Licence and attribution

The converted bundles inherit CC-BY-4.0 from the upstream model and require attribution to Kyutai. The conversion scripts and Swift host are Apache-2.0, in the GitHub repository linked above.