# Continuous server narration · 0.12.0

LLM Voice (Dev) streams first-verse MPEG audio as the provider emits it. Supported browsers append verse clips to one MediaSource sequence, preserving the media element, pitch-correct reading speed and the mixer connection across boundaries. The artificial inter-verse delay and per-verse reload are removed. Up to three following verses are prepared with two background requests at a time, within the current range/scope. Each clip has its own media timeline boundary; the visible verse follows actual playback position. Within-verse progress remains an estimate, because the provider supplies no word timestamps.

Automatic handoffs retain the stream. Manual navigation, edition changes, stop and bedtime expiration cancel old requests and queued audio. Pause retains the exact media position, including when the next verse is already buffered. A paused passage replacement prepares the new verse without starting it. Consumed audio is pruned from the playback buffer and narrator queue. Browsers without MPEG MediaSource use the existing native player, with lookahead and no added verse delay; continuous concatenation is unavailable there. Default voice remains the default, and all seven bundled texts, appearance, mixer and timer settings remain unchanged.

Streaming reduces the initial wait; it cannot eliminate provider/network delay. If generation cannot keep up, playback buffers on the current passage and never skips text. Existing caching, source verification, bounded retry, daily cap and fallback remain active. Offline exports include the continuous-player module automatically.

# LLM voice reliability audit · 0.11.3

LLM Voice (Dev) now requests the visible verse before preloading the next one. Audio initialization starts immediately; cached audio no longer depends on a preliminary server-status request. Previously heard MP3 clips persist in browser storage when supported, bounded to 128 clips and 32 MiB; the in-memory cache remains bounded separately. Playback rate reuses the same audio and preserves pitch. One bounded retry recovers temporary network/provider failures; quota and source errors are not retried. Cold generation still depends on server capacity.

The native player resumes its AudioContext before waiting for media permission, guards asynchronous startup/resume against passage changes, preserves pause during startup, and records completion arriving while paused. Startup/stall watchdogs recover unresponsive playback. Server errors retain their explanation, and fallback respects paused reading. Server narration uses the canonical text even when local pronunciation notes exist. Default voice remains the default. Mixer, bedtime, scenery, bundled text and voice labels are unchanged.

# Voice reliability update · 0.11.1

Default voice now uses installed device speech. Existing profiles switch once to this default. Server narration is labelled LLM Voice (Dev), with bounded network/audio startup and automatic fallback to Default voice on failure, retaining the visible verse. The local model remains a separately labelled development option. Mixer, bedtime, appearance and reading settings are preserved.

# The HOLY Bible - Scripture Reader · 0.11.0

Integrated from the supplied `LivingScripture_v0.8.0.html`. The reader retains seven complete bundled editions (ASV, BSB, DBY, KJV, WBS, WEB Updated and YLT), their publisher metadata, source formatting and available red-letter annotations. Scripture data is unchanged. Public-domain status and provenance are retained per edition in Versions; the importer validates structure, not the theological interpretation or independent accuracy of every word.

Open through the portfolio Tools catalogue or `index.html`. Search by reference or phrase, choose a chapter, save verses, select a reading range and customise the page and scenery. Keyboard shortcuts and the page-only view remain available. Settings use the original browser storage keys so existing compatible saves continue to load.

Local browser narration uses **George / bm_george**, British English, at 0.92× with punctuation-aware pauses and no verse-reference announcements. Voice settings offer ten alternatives, a current-verse preview, volume, speed, sentence pauses and optional speech-only pronunciation notes. These are synthetic Kokoro voices, not David Suchet or an imitation of his performance. His official NIV UK recording opens separately on Bible Gateway. Authorised local audio can be played without verse synchronisation.

Choose Listen, then enable the natural voice in Voice settings. The explicit download action fetches the model and selected voice assets from Hugging Face and the ONNX runtime from jsDelivr. The Q8 model is about 92 MB; full precision is about 326 MB. Inference runs in a module Web Worker on the device; passage text is not uploaded. The pinned kokoro-js 1.2.1 browser bundle is included locally under the Apache 2.0 licence (see vendor/KOKORO-LICENSE.txt). It includes Transformers.js 3.5.1, ONNX Runtime Web and the phonemizer. The model card is https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX and the upstream code is https://github.com/hexgrad/kokoro.

Reading and device voices work offline after the portfolio package is cached (an installed local English device voice is required for offline device speech). Neural model and voice files may persist in browser caches, but browser eviction or a new voice can require another connection. Neural offline availability is not guaranteed. Model setup is cancellable; playback errors stop at the current passage without silently skipping it. Pause/resume preserves neural audio position. Device speech resumes at the last reported boundary or replays the current phrase if the engine reports none.

Save offline reader creates a standalone HTML file containing the application, scenery, installed exportable editions, settings and pinned JavaScript voice bundle. The neural model/runtime downloads are not embedded. NIV and other licensed imports cannot be exported. Settings and bookmarks can be backed up independently. The public static site has no publisher API proxy or API-key field; official NIV reading and audio links work externally.

The portfolio build generates the launcher, case study, sitemap, offline entries and integrity manifests. Regression tests live in `tests/scripture-*.test.cjs` at repository root.

## Quiet room, mixer and bedtime

The default view tucks the full navigation and transport bars away. The small dock keeps Listen, Mixer, a moon/timer button and Controls available on touch and desktop. Controls reveals the complete interface; Escape and F8 also recover it. Drawers keep keyboard focus inside while open and return focus to the invoking control. Page settings include optional gentle light, existing scenery, motion, wind and rendering quality. Reduced-motion preferences are honoured on startup, including saved profiles.

The Sound tab mixes original synthesized piano, flute, warm tones and a procedural stereo rain texture. There are no audio recordings, additional downloads or remote services. Four presets adjust the channel levels; Play background is always explicit. Master background level and narration volume are separate. The optional automatic ducking lowers the background to 32% of its chosen level while narration is speaking/buffering, and restores it when narration pauses. Audio gain changes are smoothed. Hidden-tab sound can be disabled; it then stops without resuming itself. Page exit stops all audio and clears scheduling. Reloading keeps chosen levels and leaves all playback paused.

The moon button opens a timer of 1–240 minutes, with 15/30/45/60-minute shortcuts and a final 0/30/60/120-second fade. The deadline uses the wall clock rather than counting callbacks, persists within the current tab across reloads, and is checked on visibility changes. It keeps counting if playback is paused. Cancel restores selected volumes without starting anything. Expiry stops narration, local recordings, background sound, scheduled music and automatic reading, freezes scenery and dims the page. A deliberate new playback/Controls action wakes the page. Active timer state is excluded from preference backups and standalone exports. Neural speech, local recordings and the mixer support continuous fading; device speech volume updates depend on its browser engine and can take effect at phrase boundaries. Every engine stops at the deadline. A browser/device that suspends the entire page may delay execution until it resumes; this is not an operating-system sleep alarm.

## Fast server narration (default)

Listen now connects automatically to the deployed GPU service at https://sheedynotes-scripture-voice.pages.dev and requests a compressed MP3 for the visible verse. The default server voice is Draco (British male); Pandora, Hyperion and Theia are alternatives. These are different voices from local Kokoro George. Browser and Device choices remain available. No browser model download is needed for server speech.

The next verse preloads during playback. Advancing normally preserves this work; stop, passage replacement and expired bedtime still invalidate old playback. The screen centres and highlights the spoken verse, remains fixed while it is buffering/paused, advances only after its audio finishes, and rejects stale progress/completion callbacks. Native media playback preserves voice pitch when changing speed and keeps the exact position across pause/resume. Progress within a long verse is estimated from elapsed audio; this does not claim word timestamps.

Requests contain a bundled reference, phrase number, voice and source-text hash. The server resolves the public-domain text itself and verifies the hash before inference. Arbitrary text, private imports and licensed editions are not sent for server synthesis; choose Browser voice for imported text or speech-only aliases. No API key is shipped to visitors. Model inference uses Cloudflare Workers AI Aura 2. Audio is cached in D1, at the edge and in the active reader. Cached clips can be reused at different playback speeds.

The service limits fresh generation to 12,000 characters per UTC day, shared across users, and caps uncached requests per client. That is a maximum of US$0.36/day of model input at the documented US$0.03/1,000-character rate, before free allocation; no paid plan was enabled. Cloudflare's own account/free limits can be stricter. Cached audio remains usable after this application budget is exhausted. Service setup, quotas, repeatable deployment and source validation are documented in server/scripture/README.md in the full repository. Standalone HTML exports use this server when online; local/browser/device options remain available separately.
