BIBLE RESEARCH NEURAL AUDIO

Runtime: Microsoft ONNX Runtime Web 1.21.0.
Package: https://registry.npmjs.org/onnxruntime-web/-/onnxruntime-web-1.21.0.tgz
Source: https://github.com/microsoft/onnxruntime/tree/v1.21.0
License: MIT. See LICENSE-onnxruntime.txt and ThirdPartyNotices.txt.
The three runtime JS/WASM files are copied without modification.
Single-threaded WebAssembly inference runs in a dedicated browser worker.

Voice models: Meta Massively Multilingual Speech (MMS), VITS.
Languages: Ancient Greek (grc), English (eng), Latin (lat).
Original models:
https://huggingface.co/facebook/mms-tts-grc
https://huggingface.co/facebook/mms-tts-eng
https://huggingface.co/facebook/mms-tts-lat
License: Creative Commons Attribution-NonCommercial 4.0 International.
https://creativecommons.org/licenses/by-nc/4.0/
Research: https://arxiv.org/abs/2305.13516

ONNX exports by willwade, compatible with sherpa-onnx:
https://huggingface.co/willwade/mms-tts-multilingual-models-onnx
Pinned revision: 4f49fc254f11ddb0e35fbf25b6e48b04463a6779
Files: {grc,eng,lat}/model.onnx and tokens.txt.
Model file sizes: grc 114079236, eng 114017028, lat 114010116 bytes.
SHA-256:
grc c4787e5d7e766574f01c7f83f0d305679f60cd73ff4b1f19d33a51981e290671
eng 409bafdb550948dc5c0b216e21b341732ef3cefead8008b512b034f1bdb27132
lat ca5f134ed33d66c80644c98d4d1e1367d17ea2779fe4a0a3529f4ac55c6dddaf

Models download only when a reading button is pressed.
Completed model downloads are saved in browser CacheStorage when allowed.
A browser storage limit can prevent caching; playback can still use the download.
Stopping a download or generation terminates its worker. Completed cached
files remain available. Only one language model is held in memory at a time.
No verse, note, or word is sent to a remote speech-generation service.
Speech is synthesized locally. It is not a human recording or a verified
reconstruction of an ancient pronunciation system.

Audio-only input normalization:
- Lowercase, NFC normalization, curly apostrophe to ASCII apostrophe,
  collapse whitespace.
- English word meanings omit leading bracketed verse-number artifacts from
  STEP glosses for speech only, such as {7}, [19], or (8b).
- Greek rare characters unsupported by the model: ᾴ to ᾳ, ὂ to ὀ, ῢ to ϋ.
- Latin ligatures æ to ae and œ to oe; j to i; ë to e.
- Latin k to c is a spelling approximation for the single source occurrence
  Eliakim (Luke 3:30). It is not a correction to the written source.
- Model-supported letters retain their token IDs. Unrecognized letters,
  combining marks, or numbers cause a visible error rather than silent removal.
- Unencoded punctuation is omitted, following the model's character input.
- Token id 0 is interspersed according to the official VITS tokenizer.
- Speed changes use model length_scale, not pitch-shifting audio playback.

The displayed Bible source text and research data remain unchanged.
Neural speech output is stochastic. Generated audio is not a source quotation.

Runtime file SHA-256:
ort-wasm-simd-threaded.mjs e9ba2350c370278fc90108f1514fb9ce6a4051341ab977b5b0dca7eca9e78dfa
ort-wasm-simd-threaded.wasm 06b3f98e5aa2fffec1e3ac57a48bf1073828c6624e14d210750bc596c2e35d65
ort.wasm.min.mjs 5118f6b62b123b74a7af80dccd6dc96d3651a39d02b37edfc0d9831982f1f898