Changelog

Hotfix — Long Voice Generation Sessions No Longer Lost to “Network Error”

• Long-form voice generation now completes reliably even if the page loses connection — Previously, a long voice synthesis (multi-paragraph reflections, sermon-length input) could spend a minute or two generating audio successfully, only to surface a “network error” right at the end and mark the job as failed. The audio had actually been produced — we just couldn't tell the browser about it once the live connection had dropped (a flaky moment in the network, a tab switch, a proxy timeout). The generation flow now completes its save-to-history work regardless of whether the browser is still listening, and the recoverable-jobs system serves up the finished audio when the user comes back.
• Completed jobs stay completed — A subtle related issue: if the connection dropped at exactly the moment between “saving to history” and “sending the final “done” event”, the system would briefly mark the job as complete in the database and then immediately re-mark it as failed when the error handler ran. The user would see a failed job they couldn't recover, even though the MP3 was safely saved. Fixed by tracking completion state explicitly — the failure handler now refuses to overwrite an already-completed job.
• What you'll see differently as an editor — Mostly nothing, because the failure case was a class of intermittent error you weren't generating on purpose. If you previously hit one of these mid-generation drops, you would have lost the work and had to re-trigger from scratch. From now on, you can also recover via the jobs API (Job History pane) if the page glitches — the audio survives even if the visible progress bar dies.
• Two related code paths flagged for the same fix later — Voice creation (Voice Clone Studio upload flow) and character reference image generation (Vignettes) carry the same code pattern that caused this. Marked in code for the same defensive treatment in a follow-up release. Neither has been observed failing this way in practice, but the latent vulnerability is identified.

New Voice Synthesis Engine — Better Quality, Editorial Tone Control, Smoother Captions

• All existing voices preserved across the migration — Every voice ID in the catalogue continues to work. Authors keep the same voice settings; the editor keeps the same dropdown picker. The voice you've been hearing for a given author is still that voice — it now comes from a more capable engine that reproduces the source recording with higher fidelity. No re-saving of preferences, no re-uploading of clones.
• New British female voice: Isabella — A long-loved voice that was missing from the picker (the system tried to use a name that wasn't in our voice catalogue) has been re-captured and added. Reflection videos and Vignettes can now use Isabella alongside Alice, Emma, and Lily for British female narration.
• Editorial tone tags for the Voice Clone Studio — A new toolbar of expression tags lets the editorial team steer the prosody of any synthesis. Drop [warm] before a passage of comfort, [serious] before a moral lesson, [reverent] before a scripture quotation, [gentle] for a child-directed line. Thirteen tags total covering warm, gentle, soft, cheerful, excited, thoughtful, serious, solemn, reverent, sad, cold, firm, urgent. The chosen tone applies to the whole rendered segment.
• Three named pause buttons replace the four numeric ones — [pause-short], [breath-pause], [pause-long] instead of 0.5s / 1s / 2s / 5s. Cleaner mental model — editors think in terms of natural conversational pauses rather than millisecond budgets. This change also fixes a quiet bug: the previous numeric buttons inserted markup that the new engine couldn't parse, so the literal text “pause 2 slash” would have read aloud. The new tags work correctly.
• Reflection video captions feel smoother — Previously, very short caption lines (one, two, or three words) could flash on screen for a fraction of a second between longer captions, which felt jarring. Two complementary fixes: short sentences are now merged with adjacent ones into a properly-readable phrase (minimum four words per caption), and consecutive captions bridge their transitions so there's no visible blank-screen gap as one swaps for the next. Net result: every caption stays on screen long enough to read comfortably, and the visual flow between captions is seamless.
• Real-time progress for long audio generation — Long-form text-to-speech (sermon-length input, multi-minute reflections) now shows live chunk-by-chunk progress: “Synthesising chunk 3 of 8…” → “Synthesised 127 seconds of audio — finalising…” → “Saving to history…” instead of the progress bar appearing stuck at 20% for several minutes.
• Auto-trim for long voice uploads — When an editor uploads a voice clone reference recording longer than 15 seconds, the system now automatically extracts a centred 10-second slice (the prosody-rich middle of the recording) instead of rejecting the upload. The editorial team no longer has to manually trim recordings in an external tool before uploading.
• Auto-heal voice synthesis errors — If a voice on the new engine appears missing for any reason (rare, can happen after maintenance), the system now automatically re-registers the voice from the original source audio and retries synthesis once before surfacing an error. Editors no longer hit hard “voice unavailable” errors that would previously have required admin intervention.
• Cleaner error messages across the synthesis flow — When something goes wrong, editors now see a generic “Selected voice is unavailable. Try a different voice, or contact your administrator.” instead of the previous error which leaked internal voice IDs and repository paths into the page. Server-side logs still carry the full diagnostic for ops debugging.
• Audio analysis services consolidated to a dedicated machine — Speech recognition, forced word alignment (used for Reflection caption timing), speaker diarisation, and Arabic/Farsi text-to-speech all now run on one well-spec'd dedicated machine instead of being spread across legacy infrastructure. Faster startup, fewer cross-network hops, simpler operations. No user-visible workflow change.
• Reflections + Vignettes simplified — Voice synthesis for Reflection video shorts and Vignettes now talks directly to the new engine instead of routing through an intermediate service. One less moving part, one fewer failure mode.
• System Status & Health pages updated — The admin System Status and LLM Health pages now show the new Voice Synthesis Engine alongside the consolidated audio analysis services. The previous voice synthesis services are still listed but tagged as “(Legacy)” during the post-cutover rollback window. Cloud LLM provider tier labels were also clarified: the on-prem AI is the actual primary, with cloud providers labelled Secondary and Tertiary in the order they're consulted.
• Behind the scenes — new shared voice synthesis library — All voice synthesis callsites across the website (audio preview, article TTS, Reflections, Vignettes, Voice Clone Studio) now route through a single shared library instead of bespoke code per feature. Single source of truth for voice catalogue, markup parsing, chunk handling, and error recovery. Future voice features start from a much smaller surface.
• Routine dependency maintenance — Third-party library updates rolled in alongside this release.
• Rollback insurance — The previous voice infrastructure is being shut down (not deleted) post-deploy as a rollback safety net. If anything regresses we can flip back in minutes. Full decommission follows after the new stack proves stable in production for some weeks.

Hotfix — Vignette Art Style Consistency Across the Whole Comic

• Vignette panels now hold their chosen art style consistently across the entire comic — A subtle issue caught the day after v2.7.0 went live: vignette panels that featured a specific character (rendered through one image-generation path) were honouring the chosen art style faithfully — pencil sketch stayed pencil sketch, watercolour stayed watercolour. But scenic establishing shots between character panels (rendered through a separate, faster image-generation path) could drift into full-colour photorealism on photo-friendly subjects like architectural courtyards or modern objects. Result: a single comic could read as if two different artists had worked on alternating panels.
• Why it happened — The faster scenic-shot path was speed-calibrated in a way that didn't leave the model enough iteration budget for stylised art instructions to fully take hold. The fix moves these scenic shots to the same higher-quality path the character panels already use. Trade-off: roughly an extra five to seven seconds per scenic panel — about a minute and a quarter total for a typical ten- to twelve-panel vignette — in exchange for editorial coherence across the whole comic.
• Speed-optimised path retained where it actually fits — Reflection wallpaper batches (35 wallpapers per reflection across 7 styles × 5 ratios) keep using the fast path, since per-wallpaper style variance within a batch is invisible to the user. Character reference portraits also keep using the fast path — they're intentionally photographic anyway.
• No user-visible workflow change — Authors don't need to do anything differently. Pick a style as before; it now applies consistently across every panel of the resulting comic.

New Image Generation Engine — Better Faces, Better Hands, Bigger Vignettes

• Vignette characters look like themselves across panels — The Creative Studio now uses a new image generation engine designed specifically for keeping a character's face and identity consistent across multiple scenes. The previous engine relied on a face-transfer technique that was inconsistent across roughly half of vignette panels — resulting in subtle facial drift between scenes within the same story. The new engine works directly from the character's reference photo, preserving their face, beard, headdress, clothing details and other identifying features as they move from scene to scene.
• Hands and fingers actually look right — A persistent issue with the previous engine was hands rendering with extra fingers, fused digits, or impossible poses about one panel in five. The new engine has meaningfully better anatomical fidelity at the same hardware budget.
• Vignettes can now have up to 12 scenes for any audience — Previously the maximum scene count varied by audience style (children: 8, youth/adults: 7, seniors: 6). Now all four audience styles support up to 12 scenes, giving authors more room to tell richer stories regardless of who they're writing for.
• Art styles hold across the whole image — A subtle issue we caught during editorial testing: when an author selects a hand-drawn art style (pencil sketch, watercolour, ink wash), the engine could occasionally render the foreground character in that style but slip into photorealism for backgrounds — especially when scenes contained tech objects (phones, screens) or modern interiors. We've tightened the prompt construction so the chosen style applies consistently across the whole frame, with explicit rejection of competing styles. Pencil sketch stays pencil sketch from corner to corner.
• Faster batch wallpaper generation — Reflection wallpaper batches (35 wallpapers per reflection across 7 styles × 5 ratios) now use a faster Lightning-mode renderer that produces nearly the same quality in roughly a tenth of the time. Practical impact: a full reflection's wallpaper set finishes in under 20 minutes instead of nearly two hours.
• Title cards still render text crisply — The text rendering engine is the same model family used for the rest of the artwork, with the same legibility benefits. Common English titles render perfectly. For subtitles or attribution lines that include rare or borrowed words, the existing FFmpeg overlay layer continues to handle text precisely — a hybrid approach that gives us the best of both.
• Long-running tasks stay responsive — Vignette story generation, reflection artwork, and reflection title cards now stream live progress updates as they work. You'll see “Searching scripture for relevant passages…”, “Generating story…”, “Rendering artwork…” etc. instead of an opaque spinner. Connections also stay warm via heartbeats during longer renders, so very rich generations no longer time out before they finish.
• Reflection narration audio simplified and made more reliable — The text-to-speech path used by Reflection video shorts has been streamlined. Where it previously hopped through an intermediate service, narration audio is now generated directly. This removes a class of intermittent “connection failed” errors and matches the same approach already used by Vignette video narration — one less moving part, more consistent reliability.
• Full scene descriptions visible in the editor — The Scene Breakdown card in the vignette editor now shows the complete scene description (the actual prompt used to generate the panel artwork) rather than just a truncated header. Useful when you're reviewing or fine-tuning what the AI is going to render.
• Hardware reorganisation — The image generation service moved to a different machine on the cluster. The previous host stays warm for 1–2 months as a safety net in case anything regresses; afterwards it joins our local AI inference pool. No downtime expected.
• Faster local-AI responses — A small but pervasive optimisation: the website now explicitly tells the local AI to skip its internal “thinking” step for tasks where it doesn't add value (creative generation, title rewriting, scene authoring). Average response times for these tasks dropped roughly 3-4× without any quality loss. Reasoning-heavy tasks (translation, structured analysis) are unaffected.
• Behind the scenes cleanup — The image-generation client code was rewritten from 863 lines of imperative builders to 388 lines of clean, JSON-template-driven code. Single source of truth: the validated workflow templates we tested are exactly what runs in production. Future tunings (e.g. style adjustments) happen in one place.

Smarter Local AI & Bigger Context Window

• Local AI model upgraded to a newer generation — Our on-prem AI engine has been upgraded to a more capable model from the same family. Improvements are most noticeable in instruction following, structured output (translation, article drafts), and long-form reasoning. The same hardware serves it; the upgrade is entirely behind-the-scenes from a user perspective.
• Context window expanded 8× — The amount of source material the AI can consider in a single request grew significantly. Practical impact: longer talks can be summarised or translated without chunking, scripture-study queries can pull in more reference passages at once, and article generation has more room to think through long transcripts. Future work on the IlmFlow and FiqhFlow knowledge bases benefits from this directly.
• Same response quality, no setup changes for you — Existing AI-generated content, saved settings, and stored search vectors all remain valid; we deliberately preserved the boundary so this is a drop-in capability bump. No re-indexing, no re-saving of preferences.
• Admin: prepaid plan balance display fixed — The cloud-provider balance card on the admin Health dashboard was showing a misleading “low balance” warning for a prepaid plan that has no balance to top up. The card now shows a clear “Prepaid” label with a link to view actual usage on the provider's admin profile, instead of the false alarm.
• ScribeFlow fallback labels updated — Three ScribeFlow code paths (translation, batch translation, article generation) had model-name labels that were stale after today's upgrade; updated to reflect the current generation.

Cleanup & Simpler Architecture

• Returned to our proven single-server local AI — A short-lived experiment with a multi-machine local AI cluster (introduced in v2.6.0) did not prove reliable enough for day-to-day production use. We've rolled back to the same single-server local AI primary that served us well before, with cloud providers continuing as resilience fallback. Users will not notice any change in capability or quality — everything that worked before still works, served by the same hardware that ran our scripture translation and knowledge-base inference for months prior.
• Admin dashboards simplified to match — The system status, LLM usage, LLM settings, and health pages no longer reference the retired cluster. Historical usage rows are preserved for auditing but tagged so editors can see at a glance which entries pre-date the rollback.
• Two latent bugs fixed along the way — Found and fixed two ScribeFlow code paths that would have tried to call the retired cluster as a last-resort fallback for projects without an explicit AI configuration. Both now correctly fall back to the single-server local AI.
• Transcription path verified end-to-end — Speech-to-text and audio processing live on a different machine entirely and were never affected by the cluster work. We re-verified the full path (browser → orchestrator → speech recognition → output) after the cleanup to confirm nothing slipped.
• Lessons documented for posterity — A full post-mortem and a path-forward sprint plan have been added to internal docs so the next person to consider distributed inference at our scale doesn't have to relearn the same lessons. A planned upgrade to a newer / more efficient version of the local AI model (with richer context window) is queued as the next inference improvement.

Operations Panel Honest About Local Infrastructure

• Live health indicator on every admin page that depends on local AI — A compact green / amber / red pill in the top-right of each relevant admin page tells editors at a glance whether the home-lab is ready for a generation run. Hover or click for a per-service breakdown with response times. Polls every 30 seconds; pauses automatically when the tab is hidden.
• New LLM Health dashboard — A dedicated page at /admin/llm-settings/health shows tiles for every on-prem service (primary LLM chain, speech recognition, voice synthesis, image generation) plus remaining credits on paid cloud providers, with a low-balance warning so we top up before running large batches.
• Local AI models selectable by name in the admin UI — The LLM Settings model picker now shows our local models at the top of the dropdown (free, recommended) before the paid cloud options. Selecting a local model no longer requires configuring a cloud API key.
• Honest provider logging across every call — Every AI call on the site — transcription, translation, generation — now records which upstream actually served it. When a translation starts on our local cluster and falls back to cloud, both legs show up as distinct rows in the usage dashboard so cost and speed comparisons are accurate.
• Transcription now prefers local AI by default — The “use local transcription” toggle in ScribeFlow settings is now correctly persisted and honoured as the per-project default. Editors still have the option to override per-project when the per-project selector lands (next release).
• Translation is now local-first too — Arabic and Farsi scripture translation, and ScribeFlow SRT batch translation, now try local AI first with cloud providers as automatic resilience fallbacks. Our companion knowledge-base project has been running the same model on the same hardware for scripture translation for months at production quality — no reason for this site to keep paying cloud rates.
• IlmFlow and FiqhFlow admin panes disabled — IlmFlow has been extracted into its own standalone project; FiqhFlow is next. The admin panes for both are preserved read-only for reference but every interactive element is disabled so the tp-website database doesn't diverge from the soon-to-be-standalone repos during migration.
• Administrative email addresses no longer appear in page markup — Removed a footer that leaked super-admin email addresses into the browser DOM on one settings page. Audit history is retained server-side and viewable on the audit-logs page.
• Bug fixes — LLM settings page no longer throws “failed to fetch settings” after the database schema alignment; header pill and action button no longer wrap to multiple lines on pages with long descriptions; local AI transcription toggle now actually saves.

New Local AI Cluster — Bigger, Fresher Model, Lower Cost

• New local-inference cluster now running alongside our existing primary — Several computers previously sitting idle have been pooled into a single distributed AI cluster using their high-speed native interconnect. Together they host a newer, larger language model than our previous primary could fit, with near-lossless quality. Requests are served behind the same OpenAI-compatible API shape, so migrating any given feature across is a config change rather than a rewrite.
• Newer reference model — The cluster is running a 35-billion-parameter mixture-of-experts model released earlier this month (April 2026). It's multilingual, handles Arabic and Farsi nuance well, and is sized to our RAG / chat / article-generation workload. On real prompts the outputs we've compared are noticeably more coherent on complex multi-sentence reasoning than the previous 27-billion-parameter primary.
• Paired embedding service — The same multilingual embedding model we've been using for IlmFlow, FiqhFlow, and Reflections search is now served from the new cluster's front-door host. Bit-compatible with the vectors already in our databases — no re-embedding required.
• Frees existing primary for specialist work — Once the new cluster passes dev validation and takes over primary LLM traffic, the existing primary server will be freed up for specialist AI workloads (video generation, voice-model training, multimodal experiments). Each piece of our hardware will do what it's best at.
• Cloud fallback unchanged — Cloud AI providers remain wired as tier-2 / tier-3 fallbacks, so if the local cluster ever goes down, user-facing requests still complete without error. Users won't notice the switch; only our operations dashboard will.
• Currently in dev validation — Live on the dev site now while we compare quality and throughput against the existing primary on real workloads. Production cutover follows after 24-72 hours of validation; you'll see a follow-up changelog entry then.

Reflections Video Short — Every Word in the Reflection Now Shows in Captions

• Verse references, years, and pure-punctuation tokens are back — When the caption-timing step can't process a word (typically because it has no letters — e.g. “21:16)”, “2006,”, an exclamation-mark quote), it used to silently drop the word. The video shorts pipeline now checks what the timing step returned against the original reflection text and splices back any missing words with an interpolated display time. End result: captions carry the reflection verbatim.
• Narration and phrase timing unchanged — Narration is still the original text. Captions are still grouped into sentence-length phrases with start/end anchored to confidently-timed words, so the approximate timing on spliced words is absorbed into the phrase's display window.

Reflections Video Short — Verse References Now Appear in Captions

• Verse references like “(Quran 21:16)” now show up in captions in full — v2.5.8 restored punctuation and case to the caption-timing input, but words that consist entirely of non-letter characters (a chapter:verse reference, a bare year like “2006.”, etc.) were still being dropped because of a second layer in the same internal rebuild step. The caption-timing input now correctly keeps such words through to the output.
• Narration unchanged — Purely an internal caption-text-preservation fix. What gets spoken is unchanged.
• Better internal diagnostics — The caption-timing service now logs a small sample of its tokenised input on every request, so if any similar class of character drop happens in future we can see the cause in the first log rather than the third test run.

Reflections Video Short — Full Caption Fidelity Restored

• Verse references, quoted citations, digits and punctuation now appear in captions — Earlier releases (2.5.5 → 2.5.7) progressively stripped characters from the caption-timing input as part of fixing a separate alignment crash, and those stripped characters were silently disappearing from the on-screen captions too. Now that the underlying alignment fix is in place at the service level, the caption-timing input passes through unchanged, and text like “(Quran 21:16)” shows up in captions exactly as written.
• Mixed case preserved — Title-case phrasings (“Creating Heaven and Hell Within Ourselves”) and proper nouns render in their original case rather than being flattened to lowercase.
• Narration unchanged — As in every previous release of this work, narration still uses the original text; no change to what is spoken or how it is paced.

Reflections Video Short — Root-Cause Fix for Alignment Failures

• Video shorts reliably again — The last two releases tried to clean the input text more aggressively, but the real bug was on the caption-timing side. An internal step that splits long tokens back into individual letters was also turning an optional “skippable” marker into a hard word boundary — and when the audio had actual sound at those positions, the timing engine would reject the whole request. That step now leaves the skippable marker alone, and the timing engine completes cleanly.
• Narration and captions unchanged — No change to what gets spoken or shown on-screen. Purely internal plumbing.
• Deep diagnostics baked in — If anything of this class ever breaks again, the server now logs the exact internal state at the point of failure (lengths, sample contents, where the two sequences diverged), so the next fix can land the same day without guesswork.

Reflections Video Short — Stronger Text Cleanup for Scripture-Heavy Content

• Video shorts now handle Quranic quotations and verse references cleanly — The v2.5.5 pass caught smart quotes and accented transliterations but still left straight double quotes and verse references like “21:16” in the caption-timing input, which the aligner couldn't parse and which triggered the same “Generate Video Short” error on reflections that quoted scripture. The caption-timing step now uses a strict letters-and-spaces only pass, so any punctuation, digits, or symbols in the reflection text are stripped cleanly before captions are timed.
• Narration is unchanged — The stricter cleanup only applies to the caption-timing step. Your reflection's full text — including quotes, verse numbers, and punctuation — is still spoken exactly as written.
• Better diagnostics if a new edge-case appears — If a future reflection ever trips a similar error, the server now logs the specific characters involved so the fix can land the same day rather than needing the original reflection to be reproduced locally.

Reflections Video Short Reliability, Safer Generation Flow & Voice Clone Studio Fix

• Video shorts for Reflections now cope with fancy punctuation — Smart quotes, em/en dashes, ellipsis characters, non-breaking spaces and accented transliterations occasionally tripped up the caption-timing step, causing the “Generate Video Short” action to error out. These characters are now normalised to their plain-text equivalents before captions are timed, so pasted reflection content is handled reliably.
• Accidental tab close is now guarded — While an image, wallpaper, title card or video short is being generated for a reflection, the browser now prompts for confirmation before a close, refresh or back-button takes you off the page. Long generations can take a few minutes, and leaving mid-run used to discard the result; the prompt prevents that by accident.
• Voice Clone Studio “chunk failed” error resolved — A mismatch between the voice-cloning model and its audio-watermark helper meant the very first chunk would fail with a 500 server error the moment you clicked Generate. The helper now exposes the expected method names, so voice cloning generates cleanly again.
• No change to narration — The punctuation normalisation applies only to the caption-timing input. The spoken narration still uses the original text, so pronunciation and pacing are unchanged.

Subtitle Translation Stability Fix

• Translation jobs now recover automatically — If the AI returns a response that's slightly malformed (a stray quotation mark, an awkward line break, a cut-off reply), the system now cleans it up and carries on, instead of aborting the whole job.
• One difficult segment no longer kills the batch — If recovery isn't enough, the problem batch is automatically split in half and each half is translated separately, isolating the tricky segment without losing everything else.
• Clearer error messages — On the rare occasion a translation genuinely can't be recovered, you'll see a more informative message (with which batch failed) instead of a raw technical error.
• Stricter guidance to the AI — We've tightened the instructions the translator receives, so the kinds of malformed responses that caused the original problem are less likely to happen in the first place.

Storage Moved to Dedicated Hosts

• Scripture and rulings data on their own machines — The databases that power search, study chat, and scripture suggestions have moved onto dedicated hosts optimised specifically for that workload. Faster indexes, more room to grow.
• Graphics hardware freed up — The machines that previously shared a host with the databases are now fully available for upcoming visual features we've been planning.
• Zero downtime, zero data loss — The move happened with a single container restart; every verse, every translation, every ruling transferred across intact.
• No reader-facing change — Search still searches, chat still chats, everything looks and feels the same.

Behind-the-scenes Tidy-up

• Old AI services switched off — Following last update's consolidation, the smaller AI helpers we used to run alongside the main one are now formally retired. The platform behaves identically; we've just removed the moving parts that were no longer doing anything.
• Resources freed for new features — The hardware those services used to occupy is now available for upcoming additions on our roadmap.
• No reader-facing change — Nothing to do, nothing to notice.

Smoother Vignette Creation

• Automatic second attempt — When the AI occasionally produces a story that's longer than the chosen audience length, the system now quietly asks it to try again with tighter wording before giving up. Most generations that used to fail now succeed on the retry.
• Clearer error messages — On the rare occasion both attempts still don't fit, you'll see a friendly message explaining what happened and suggesting a fix (e.g. a more focused premise), instead of a technical parse error.

AI Infrastructure Streamlined

• Unified AI assistant — The AI features powering reflections, vignettes, study chat, and scripture suggestions now all run through a single, larger, more capable model. Expect slightly sharper answers and more consistent writing across the site.
• Simpler under the hood — Several smaller AI services have been consolidated into one, reducing moving parts and making the platform easier to keep running reliably.
• No changes for readers — Everything looks and feels the same. You don't need to do anything. The upgrade is purely on the server side.

Sharing & Thumbnail Fixes

• Better vignette share previews — Sharing a vignette now uses a concise, automatically generated summary sentence instead of raw story content. Existing vignettes fall back to a story excerpt.
• Vignette archive excerpts — Vignette cards on the archive page now also display the generated summary instead of internal content.
• Correct reflection video thumbnail — The homepage reflection video now shows the title card as its thumbnail, matching the detail page behaviour.

Video Caption & Alignment Fixes

• Caption alignment fix — Fixed word-level caption timing that broke when narration text contained parentheses or brackets. These characters fall outside the alignment model's vocabulary and caused the entire caption track to fail silently. All special punctuation is now stripped before alignment.
• Accurate word timestamps — Fixed an issue where word-level timestamps were compressed to a fraction of the actual audio duration, causing captions to appear and disappear far too early. Timestamps now span the full narration correctly.
• Cleaner word boundaries — Eliminated phantom zero-length words that inflated caption word counts (e.g. 857 words instead of 177), producing cleaner subtitle tracks with accurate word-by-word highlighting.
• Snappier title card — Title card now fades out after 4 seconds (previously held for the full intro duration), with a quicker 1-second fade for a smoother transition into content.
• Full vignette pipeline verified — End-to-end vignette generation (story, images, panels, audio, captions, video) confirmed working across all audience styles.

Multilingual Audio & Consolidated Voice Services

• Consolidated audio infrastructure — Unified 4 audio/voice services onto a single high-performance server for faster processing and simpler maintenance
• Arabic & Farsi voice synthesis — New text-to-speech support for Arabic and Farsi, enabling audio generation for multilingual content
• Local AI transcription — On-premises speech recognition supporting 52 languages with dramatically better quality on noisy audio, accents, and dialects. No external API dependency
• Document audio generation — New admin tool for generating audio from pasted text with automatic language detection and voice routing
• Multilingual static page audio — Static page audio now handles Arabic and Farsi content alongside English, producing seamless merged output
• Local AI translation — Subtitle translation now uses the local high-performance LLM by default, eliminating cloud API costs with automatic fallback
• Dynamic video encoding timeout — Video composition timeout scales with duration, fixing errors on longer videos
• Real-time encoding progress — Video generation shows actual encoding percentage in the admin UI

Network Resilience — Video Generation

• Automatic retry for voice generation — Video generation now automatically retries on transient network failures with exponential backoff (up to 3 attempts), fixing intermittent errors during audio generation
• Timeout protection — All voice synthesis calls now have configurable timeouts preventing indefinite hangs during video generation
• Storage fetch retry — Asset fetches (panel images, sound effects, music tracks, artwork) retry automatically on transient failures across both Vignettes and Reflections
• Clearer error messages — Video generation errors now identify the specific scene and step that failed, replacing generic error messages

ScribeFlow — Transcription Model Selector

• Dynamic model discovery — Available transcription models are now detected automatically based on audio input capability, so new models appear without manual updates
• Searchable model selector — Replaced the basic dropdown with a searchable, filterable picker showing model details and recommended badges
• Active model indicator — Visual indicator showing the currently saved transcription model at a glance
• Broader model support — More AI transcription models now available for selection, covering a wider range of providers and capabilities

Intelligent Content Generation

• Upgraded AI content generation with a new local-first architecture for faster, more reliable results
• All content generation tasks (reflections, vignettes, artwork prompts, video metadata) now use intelligent routing with automatic failover
• New LLM Usage dashboard for administrators to monitor AI request patterns and costs
• Unified content generation pipeline — consistent quality across all features
• Reduced dependency on external AI services with local inference for ~85% of tasks
• Content is now enriched with authentic scripture references from our knowledge base — reflections and vignettes draw on real hadith, duas, and Quranic verses
• Every AI request is now audited with token counts, response times, and cost tracking

March 22, 2026

Image Generation Improvements

Editor UX

Background Audio Management

Community & Contact

SEO & Page Quality

March 21, 2026 — Eid al-Fitr / Nowruz

Vignettes — Islamic Anecdotes

A new content format alongside Reflections — short moral anecdotes rooted in Islamic scripture, inspired by classical storytelling traditions. Each vignette targets a specific audience with tailored writing, visuals, audio, and video.

Reflections Enhancements

SEO & Social Sharing

March 17, 2026

Expanded Scripture Library — 6 Sources

UX Overhaul — Consistent Headers & Navigation

Reflections Editor Enhancements

Video Shorts Improvements

Bug Fixes & Improvements

IlmFlow Study Workspace

IlmFlow Data Quality Pipeline

IlmFlow Short URL Sharing

New Domains & Pre-Release

February 14, 2026

IlmFlow — Islamic Knowledge Base

Reflections — Daily Thought & Scripture

February 3, 2026

Substack Editor App

January 17, 2026

Voice Synthesis Stability

Task Recovery System

January 16, 2026

Network Administration Dashboard

January 16, 2026

Dramatically Faster Translation

January 16, 2026

Code Quality & Maintenance

January 15, 2026

Bug Fixes

January 6, 2026

Bug Fixes

January 3, 2026

New Feature: FiqhFlow Islamic Rulings

January 2, 2026

Enhancements

January 2, 2026

Enhancements

December 31, 2025

Bug Fixes

Enhancements

December 30, 2025

Bug Fixes

Enhancements

• •Model dropdown shows all available audio-capable AI models
• •1-hour cache with automatic fallback for model list

December 28, 2025

New Features

Enhancements

• •Publication name now shown in edit modal header for context
• •Improved voice synthesis reliability with distributed processing
• •Full audit logging for all domain configuration changes

December 28, 2025

Bug Fixes

• •Fixed changelog menu showing v1.6.0 as "Unreleased" after release

December 28, 2025

Major Features

Enhancements

Bug Fixes

Released: December 27, 2025

Bug Fixes

• •Fixed Voice Clone Studio pause/expression buttons causing textarea to scroll to top
• •Fixed static page audio incorrectly marked as "outdated" after server updates
• •Fixed static page audio showing impossible count (e.g., "8/7 pages")

Version 1.5.2

Released: December 27, 2025

Bug Fixes

• •Fixed transcription engine selection causing errors due to ID format mismatch

Version 1.5.1

Released: December 27, 2025

Bug Fixes

• •Fixed transcription engine selection not being honored for per-job settings

Version 1.5.0

Released: December 26, 2025

Major Features

Enhancements

• •TTS now announces URLs: "Read more at truthpromoters.com/page-name"
• •Contact buttons include email: "Contact us at [email protected]"
• •Delete button now cleans up audio files from storage
• •Activate button properly links audio to pages
• •Full voice names displayed instead of technical IDs
• •Comprehensive audit logging for all TTS operations
• •Dashboard section renamed from "TTS" to "Audio Generation"
• •New "Audio Verify" button in admin dashboard

Bug Fixes

• •Fixed voice cloning text truncation at ~280 characters
• •Fixed pause tags not producing silence in audio output
• •Fixed dark mode visibility for paralinguistic tag buttons
• •Removed unsupported vocal effects that were being spoken as text
• •Fixed page audio incorrectly reading URLs (slashes were being spoken as "or")

Security

• •System status page now uses generic service names
• •Port numbers hidden from admin interface
• •Improved operational security across all admin pages

Version 1.4.1

Released: December 25, 2025

Bug Fixes

• •Fixed audio player not appearing on Help Center pages (e.g., Getting Started, Prayer Times)

Version 1.4.0

Released: December 25, 2025

Major Features

Enhancements

• •Users can now subscribe from multiple devices/browsers
• •City and country automatically detected from your connection
• •IP geolocation fallback when browser location is denied
• •Improved notification click handling in service worker

Bug Fixes

• •Fixed subscription error when user already subscribed on another device
• •Fixed "Location required" error when browser geolocation is blocked

Version 1.3.0

Released: December 23, 2025

Major Features

Enhancements

• •Voice isolation now handles long audio files with smooth crossfades
• •Fixed handling of very short audio segments
• •Quality tier settings: Fast, Medium, and Thorough processing modes
• •Translation and article generation now support all AI providers with fallback

Bug Fixes

• •Fixed API key validation error on settings page
• •Fixed model routing based on provider detection

Version 1.2.1

Released: December 21, 2025

Bug Fixes

• •Fixed: Super Admin features now correctly hidden from Author role users
• •Fixed: User Management link now only visible to Super Admins

Version 1.2.0

Released: December 21, 2025

Major Features

Enhancements

• •Queue status check before starting transcription
• •Immediate cancellation of operations
• •Enhanced real-time progress streaming
• •Artifact badges showing T (Transcription), Tr (Translations), A (Article)
• •Comprehensive audit logging for all operations
• •Clean markdown rendering for generated articles

Version 1.1.1

Released: December 18, 2025

Bug Fixes

• •Added automatic 500ms pauses between sentences in TTS synthesis
• •Manual pause marker support: <pause 500ms/>, <pause 1s/>, etc.
• •Fixed job progress tracking issues

Version 1.1.0

Released: December 16, 2025

Major Features

Enhancements

• •Error state recovery with dismiss button
• •Dark mode blockquote visibility fix
• •Translation anti-hallucination safeguards
• •Cancel buttons for all long-running operations
• •Per-user content filtering (users see only their own projects)

Version 1.0.2

Released: December 12, 2025

Bug Fixes

• •Fixed dropdown menu visibility in article actions
• •Fixed media downloads during article ingestion
• •Fixed custom TTS voice creation
• •Fixed article image display issues

Version 1.0.1

Released: December 12, 2025

Bug Fixes

• •Fixed file upload issues in various admin sections
• •Improved environment isolation between development and production