TaleChime runtime resources
Public, ungated runtime data for TaleChime and CastGlean. Downloading does not require an account or access token. Consumers must pin a commit and verify file sizes and SHA-256 against their approved resource recipe.
Kokoro Mandarin frontend v1
kokoro/misaki-zh-v1/ contains dictionaries, segmentation/POS-HMM data and
pronunciation rules exported from pinned Misaki 0.9.4. It is used by the native
Rust frontend while acoustic inference stays on Candle. It contains no acoustic
ONNX model, Python runtime, application credentials or user audio.
The package includes a provenance/checksum manifest, independent test corpus, source notices and complete Apache-2.0/MIT license texts. See its NOTICE and manifest for exact source revisions, dependency versions and validation boundaries.
Frontend validation: 28 Mandarin cases match pinned Misaki output; one mixed English case is explicitly unsupported. Acoustic, device and listening acceptance are separate. This resource publication is not an application or SDK release.
Original model weights and voice packs are obtained from their separately pinned upstream repositories; their terms remain applicable.
Optional g2pW context tables v1
kokoro/g2pw-context-v1/ contains fixed pronunciation tables and provenance for
the optional contextual frontend. The acoustic engine remains Candle; ORT CPU
is used only by the separately downloaded official g2pW pronunciation model.
The official ONNX model/archive and BERT vocabulary are obtained directly from
their pinned upstream repositories. This package does not republish those weights.
All 24 independent official input/output cases match the native runtime, including tokenizer inputs, masks and output probabilities. Manual target gold is 23/24 for both official and native execution; see the package NOTICE and manifest. Apache-2.0 terms are included in full.
Verified runtime components v3
native/ contains only the runtime components that required native conversion or assembly. Each subdirectory includes a provenance record, exact hashes and component-specific license notices. This repository is a mixed-license resource collection: IndexTTS weights retain the bilibili Model Use License Agreement and are not licensed under Apache-2.0. Users of those weights must comply with its included terms and restrictions.
- Kokoro: F32 folded Candle weights and model manifests; the optional g2pW pronunciation ONNX is extracted unchanged from its fixed upstream archive. No ONNX acoustic model is used.
- CosyVoice3: HiFT with F64 F0 tensors, complete tokenizer and fixed configuration.
- IndexTTS: GPT, codec, s2mel, CAMPPlus, BigVGAN and feature tensors in verified F32 layouts, with configuration and provenance.
Unchanged source files stay in their pinned upstream repositories. Conversion is performed by the maintainer; application users download the approved finished files and verify hashes. Native build recipes remain reproducible. Resource publication does not imply successful synthesis, listening acceptance, an SDK release or an application release.