Modelpublic

theoracleguy/pocket-tts-coreml-v3

sourceHugging Facecc-by-4.0updated 14d agoView on Hugging Face

0likes

Model Card

Pocket TTS Core ML v3 runtime

Runtime-only, non-24L Pocket TTS packages for iOS/macOS.

Source: FluidInference/pocket-tts-coreml at commit 91748676fe3c8b2eb3007b3125253bcd898202c3.

Included for every language:

—cond_prefill_ane.mlmodelc
—flowlm_step_ane.mlmodelc
—flow_decoder_fused.mlmodelc
—mimi_decoder.mlmodelc
—complete constants_bin tokenizer, embeddings, protocol constants, and voices
—a per-language download manifest with file sizes and SHA-256 checksums

German, Italian, Portuguese, and Spanish include their language-specific mimi_encoderv3.mlmodelc. English uses the upstream shared mimi_encoderv2.mlmodelc, because no English mimi_encoderv3 is published. The English pocket_state.mlmodelc is included for compatibility/testing; the rank-4 ANE path remains the stable runtime set.

The shared Mimi initialization tensors and encoder recovery matrix are also included for compatibility with existing FluidAudio hosts. The English v2 encoder is additionally retained at its established repository-root path.

The original Pocket TTS and converted weights are distributed under CC BY 4.0. See the upstream model card for attribution and base-model details.