Bible Research: English voice and pronunciation dependencies

Model: KittenML/kitten-tts-nano-0.8-int8, 15 million parameters, Apache-2.0.
Pinned revision: 84781d74e29ee25217551556398b42f80593a813
https://huggingface.co/KittenML/kitten-tts-nano-0.8-int8/tree/84781d74e29ee25217551556398b42f80593a813
The model card is preserved as MODEL-README.md. Apache license text is in LICENSE-kitten.txt.
Model URL: https://huggingface.co/KittenML/kitten-tts-nano-0.8-int8/resolve/84781d74e29ee25217551556398b42f80593a813/kitten_tts_nano_v0_8.onnx
Model size: 24,369,971 bytes
Model SHA-256: f7b0afcbee92870b32b8e0276d855b954dc25470c9f051b376ac7eee537c76fc

Fixed speaker: Bruno, official voice id expr-voice-3-m (male).
bruno.bin is the unmodified little-endian float32 payload of expr-voice-3-m.npy,
extracted from voices.npz at the same model revision. The NPY header is omitted.
The embedding has 400 rows and 256 columns. No speaker blending is performed.
Source archive: https://huggingface.co/KittenML/kitten-tts-nano-0.8-int8/resolve/84781d74e29ee25217551556398b42f80593a813/voices.npz
Source archive size: 3,278,902 bytes
Source archive SHA-256: 8aa7cee235abb0739cb51e6559685f65a4dacd95568833d05699b1633f519b3f
bruno.bin size: 409,600 bytes
bruno.bin SHA-256: 388af21b8452d723738e1cc625838182cd306e36cbffc3ab2e9f310644783a18
Model and voice are downloaded or read from cache only after size and SHA-256 validation.
The model and speaker total is 24.78 MB decimal, excluding the first-use runtime.

Official KittenTTS Web source: Apache-2.0, version 0.1.0.
Pinned revision: c846fe92858f3447f4912226471f5a63983b3f94
https://github.com/KittenML/KittenTTS-web/tree/c846fe92858f3447f4912226471f5a63983b3f94
kitten-phonemizer.js is the official examples/html/assets/kittentts-web.js browser
bundle, with only an ESM export and explanatory comment appended. The original
bundled notices remain intact. The application constructs CEPhonemizer only,
using local rule text and dialect en-us. It does not construct KittenTTS or any
SDK analytics client. Text is phonemized and synthesized on the user's device.
No API key or speech service account is required.
https://github.com/KittenML/KittenTTS-web/blob/c846fe92858f3447f4912226471f5a63983b3f94/src/phonemizer/CEPhonemizer.ts

Inference follows the official TTSEngine and TextCleaner contracts:
https://github.com/KittenML/KittenTTS-web/blob/c846fe92858f3447f4912226471f5a63983b3f94/src/engine/TTSEngine.ts
https://github.com/KittenML/KittenTTS-web/blob/c846fe92858f3447f4912226471f5a63983b3f94/src/engine/TextCleaner.ts
IDs are [0, ...phoneme IDs, 10, 0]. Bruno uses style row min(body ID count,399),
and effective speed is the user speed multiplied by the official Nano prior 0.8.
Output is mono float32 PCM at 24,000 Hz. All inference tensors are disposed.
The local tokenizer.json was independently extracted from pinned TextCleaner.ts.
The helper keeps complete normalized text and recursively splits at word or
sentence boundaries until each input contains at most 400 IDs including padding.
Unsupported phoneme letters or marks fail clearly, rather than being discarded.

english-input.js is the application adapter. It verbalizes numbers and section
labels, preserves words and source punctuation for model pauses, and keeps the
source verse unchanged. Accented Latin names remain in chunk.text while the
phonemizer-only input folds Latin accents, preventing CE from dropping vowels.
Non-Latin source letters fail clearly in this English-only engine.
American rhotic correction for words such as surface follows Misaki's US fallback:
https://github.com/hexgrad/misaki/blob/fba1236595f2d2bf21d414ba6e57d25256afada3/misaki/espeak.py
Misaki revision fba1236595f2d2bf21d414ba6e57d25256afada3, Apache-2.0.
LICENSE-misaki.txt contains the license. Kitten's native syllabic U+0329 is retained.
The output uses American English pronunciation rules. This is not a claim about
the recording location or accent provenance of Bruno's original training audio.

English rules and dictionary are unmodified text source files from eSpeak NG:
Revision 59eb19938f12e30881c81d86ce4a7de25414c9f4
https://github.com/espeak-ng/espeak-ng/tree/59eb19938f12e30881c81d86ce4a7de25414c9f4/dictsource
Files en_rules and en_list retain their copyright notices. GPL-3.0-or-later,
license preserved in LICENSE-espeak.txt. These source files are distributed here
in their editable original form and loaded locally after byte/hash verification.

ONNX Runtime Web 1.26.0, Microsoft, MIT license.
https://github.com/microsoft/onnxruntime/tree/v1.26.0
Runtime files are unmodified copies of npm onnxruntime-web@1.26.0/dist:
ort.wasm.min.mjs, ort-wasm-simd-threaded.mjs, ort-wasm-simd-threaded.wasm.
Original source URL prefix: https://cdn.jsdelivr.net/npm/onnxruntime-web@1.26.0/dist/
LICENSE-ort.txt and ThirdPartyNotices-ort.txt preserve upstream notices.
Inference explicitly uses single-thread WASM and does not require WebGPU.
The Kitten browser bundle also includes pako 2.1.0 (MIT AND Zlib), with
LICENSE-pako.txt and its original bundled notice preserved.

SHA256SUMS lists all files distributed in this directory except itself.