Indic Reader

🌐 Live app: https://indic-reader.pages.dev/

A single-file web app for reading, narrating, and translating Indic-script scriptures (Sanskrit, Hindi, Gujarati). Works in any modern browser. Everything runs locally on your device by default — no server, no tracking, no cookies. Cloud-based features (OCR, translation, dictionary) are optional and opt-in with explicit consent.

Table of contents

1. Quick start
2. Loading documents
3. Reading & narration
4. Full-screen immersive reading
5. Layouts (A / B / C)
6. Language filter
7. Outline / table of contents
8. Word meanings & translation
9. Split Sandhi (word separation)
10. Correcting language & text
11. Vedabase deep links
12. OCR for scanned documents
13. Re-OCR a region
14. Cloud OCR setup
15. Google Drive integration
16. Settings & preferences
17. Keyboard shortcuts
18. Privacy & data handling
19. Troubleshooting
20. Known limitations

Quick start

1. Open index.html in any modern browser (Chrome, Firefox, Safari, Edge — desktop or mobile)
2. Drag a file onto the drop zone, or tap 📁 Choose Files
3. For PDFs, select page range, tap ▶ Process Pages
4. Use the ▶ Play button to narrate, or scroll/tap to read

For scanned PDFs (image-only), turn on Force OCR before processing.

Loading documents

Five ways to load content:

Supported file types: PDF, DOCX, TXT, RTF, HTML, XML, images (JPG, PNG, etc.)

Pasting formatted text

When you paste from a rich source (a web page, a document editor), the app reads the clipboard's HTML so paragraphs, line breaks, and bullet lists are preserved rather than flattened — bullets are normalized and each paragraph keeps its own line. A small indicator tells you whether rich formatting was detected. If a source only provides plain text (some apps strip formatting on copy), a ¶ Restore paragraphs button reconstructs structure by breaking at sentence ends and before section numbers, and recognizes a wide range of bullet markers. Pasted text starts a clean document, so it never disturbs a PDF you already have open (its reading position and OCR cache are preserved). Structure carries through to both Text View and Document View.

Reading & narration

Two views, switchable via the tabs at the top of the reader:

• 📝 Text View — extracted text reflowed for screen reading. One line/verse per row. Each row has a ★ favorite button.
• 📄 Document View — original page rendered as image with text overlay highlighting. Shows the page exactly as printed.

Player controls (bottom bar)

Narration chips (floating player)

Reading-time options sit on the floating player as compact chips that highlight when active (tap to toggle), kept on a single line:

• 🕉 Chant — slow, contemplative narration speed (overrides per-language speed)
• Scroll (Auto-scroll) — scrolls the page automatically as narration progresses, in both Text and Document view
• Pause-skip (Pause at skip) — when narration encounters a line in a language you've filtered out, pause briefly so you can decide whether to skip or include it
• 📖 Cont. (Continuous) — auto-advance through pages without manual ⏭. In Continuous mode the app OCRs the next page just-in-time: a few lines before the current page ends it prepares the next page in the background, then narration flows straight into it — you no longer have to pre-select every page. Each page prepared this way is written to the OCR cache, so it's instant (and free of repeat cloud cost) next time.

⚙️ Settings panel

Set-and-forget options live in a single ⚙️ Settings panel (button beside Library), grouped into OCR, Playback, Reading layout, and Document, so the reading surface stays uncluttered:

• 📞 Pause on call — auto-pause narration on a phone call or when the tab is hidden, resume after
• 📍 Resume — remember your position (page, line, word) and offer to continue; reopening a multi-page PDF via Continue auto-selects and processes just the page you were on, then jumps to the exact line/word
• 🗄 OCR cache — cache recognized text + word boxes per page so reopening/re-reading is instant and avoids repeat cloud-OCR cost (turning it off clears the cache)
• ⚠️ Check OCR — underline words the OCR engine was unsure about (low confidence) so they're easy to spot and fix
• 🔋 Screen-off play — keep narrating when the screen turns off or the app is backgrounded, where the device allows it (off by default; see Background & car playback below for the trade-off)
• 📄 Doc view on play — switch to Document View automatically when narration starts
• 🔊 Read translation in narration — also speak each line's translation during continuous narration
• Force OCR, ☁️ Cloud OCR, ↩ Fallback to local, ⚙ Cloud setup — OCR engine configuration
• Reading layout A / B / C — the layout selector
• 📋 Copy Filtered — copy only the lines matching the current language filter

Full-screen immersive reading

A ⛶ Full screen tab puts the reader into a distraction-free mode showing only the current view (Text or Document) with a single floating control bar — minimise ⤢, ⏮, play/pause, ⏭, a Text/Document toggle, and a language filter. The app header, upload area, settings, and all editing tools are hidden. Document View fills the entire screen in this mode (the normal half-height cap is removed). The control bar stays visible while you read, and the app enters full-screen automatically once a page is processed, so you go straight into reading. Exit any time with ⤢. On Android Chrome this also engages the browser's true full-screen; on iOS it relies on the immersive layout (and is fully chrome-free when the app is added to the Home Screen).

Per-verse translation narration

Each translated verse block carries its own 🔊 button next to the translation text — tap it to hear that verse's translation read aloud in the right voice for the target language, independent of the main narration.

Background & car playback

While narrating, the screen is held awake so playback doesn't stop when the display would otherwise sleep. With a phone connected to a car by Bluetooth, narration plays through the car speakers and the basic steering-wheel/head-unit play/pause may control it. Turning on 🔋 Screen-off play keeps a silent media session alive so lock-screen transport controls appear and playback continues when backgrounded — but because that media session competes with the speech channel, it can make narration less stable on Android Auto; leave it off for the most reliable in-car audio. A web app cannot appear as a native app on the Android Auto / CarPlay projected screen — that requires a native build — so use Bluetooth audio mode in the car.

Line numbers

Text View shows a muted line number in the left gutter of each line, to make it easy to reference a line for correction. The numbers are drawn so they're never included when you select or copy text.

Speed chips (per language)

Narration speed is set per language (Gujarati, Hindi, Sanskrit) with three compact speed chips showing the current value (e.g. ગુ 1.0×). Tap a chip to expand a fine-tune slider for that language; tap again to collapse. Sanskrit chanting is traditionally slower; Hindi/Gujarati prose can be faster. The default is 1.0× for all three (existing saved preferences are kept). The Sanskrit chip drives both plain reading speed and chant-mode pacing across its full range.

Screen wake lock

While narrating, the screen stays on automatically. When you stop narration or close the tab, it releases. Requires HTTPS; some browsers (older Safari) don't support it and screen behavior falls back to OS default.

Chhand (verse metre) & scansion

Each Sanskrit verse shows a metre pill, e.g. 📐 Śārdūlavikrīḍita · 38 akṣ. Metre is identified two ways:

• By guru/laghu scansion (rigorous): the app marks each syllable heavy (guru, –) or light (laghu, ‿) using the actual rules — long vowel, anusvāra/visarga, or a following consonant cluster make a syllable heavy — then matches the heavy/light pattern against known metres (Indravajrā, Upendravajrā, Vaṃśastha, Vasantatilakā, Mālinī, Śārdūlavikrīḍita, Sragdharā). When the pattern matches, the exact metre name is shown with no "approx." hedge.
• By syllable count (fallback): if no pattern matches, it falls back to count-based families — Anuṣṭubh (32), Gāyatrī (24) shown reliably; Triṣṭubh/Jagatī/long shown as "(approx.)".

A verse has two related, separate controls:

• 📐 Scan — opens a syllable-by-syllable grid of the whole verse with – guru (heavy) / ‿ laghu (light) above each akṣara. This is the dense, rigorous view of the metre.
• 🎵 Chant tones — shows each pāda on its own row, labelled with its recitation tone (▲ high / ◆ mid / ▼ low), with the pāda kept as readable whole words and the – / ‿ marks sitting on a small line above each word. This is the reading-friendly view: it keeps the original chant-mode text (words are never broken into a per-akṣara grid) and simply adds the scansion marks on top. Each pāda has a 🔊 to hear it, and 🎵 Chant whole verse plays them in sequence.

The tones are a recitation aid (not canonical Vedic svara — laukika verses carry no textual accent); the guru/laghu marks are the rigorous scansion. Both views share the same scansion engine, so the marks always agree.

Syllables display with their full conjunct spelling: a consonant cluster that opens a word (e.g. the स् of स्वर्ण, the क् of क्षेत्रेषु) is shown attached to the syllable it begins (स्वर्, क्षेत्), while a cluster in the middle of a word closes its syllable (वर्). This is a display refinement only — the heavy/light weights and metre detection are unchanged.

Honest limit: scansion is only as accurate as the text. OCR errors (a wrong vowel length, a dropped conjunct) shift the marks, so verify against a clean source for serious study. The metre also shapes chant-mode recitation pacing.

Layouts (A / B / C)

Top of the reader has a layout selector. Choose what suits your reading style:

• A — Single column (default): on-demand translation. Tap "Translate verse" on each line if you want it.
• B — Side-by-side: Sanskrit on the left, translations on the right. Auto-translates as lines scroll into view. Desktop-only — falls back to C below 768px width.
• C — Stacked pairs: each Sanskrit verse followed immediately by its Gujarati + Hindi + English translation. Auto-translates everything. Best for mobile.

Choice is saved in your browser; persists across sessions.

Switching to B or C the first time will prompt for consent (translation sends text to Google Translate's public endpoint).

Language filter

Buttons above the text: All / ગુ Guj / સં Sans / હિ Hin

• Filters which lines are visible AND which lines are narrated
• Useful when reading scripture with both Sanskrit verses and Gujarati commentary — you can hear just the Sanskrit, or just the commentary

Outline / table of contents

For DOCX files, the 📑 Outline button appears in the reader header. Tap it to see the document's structure as a navigable list.

The outline shows different things depending on the document:

Tap any entry to scroll the reader to that section. Brief highlight pulse shows where you landed.

Word meanings & translation

Show Meanings (Sanskrit only)

Below each Sanskrit verse, tap Show Meanings to look up each word:

1. Queries the Cologne University Monier-Williams dictionary (the authoritative Sanskrit lexicon)
2. Returns English definitions, then auto-translates them to Gujarati
3. Cached per-session — repeated lookups are instant

Source badge per word:

• MW (green) — from Monier-Williams (authoritative)
• GT (grey) — Google Translate fallback (less reliable for technical Sanskrit terms)

Translate verse

Below each line, tap 🌐 Translate verse to get full-verse translations:

• Gujarati, Hindi, and English translations all shown
• Each has a 🔊 button to hear it spoken in that language's voice
• Cached per source — same verse won't be translated twice

This sends the verse text to Google Translate's public endpoint. Disabled by default; first use prompts for consent.

Translate page (Gujarati ↔ Hindi)

In Text View, three buttons in the controls row translate the current page, always leaving Sanskrit verses untouched:

• 🌐 ગુ→हि — translate the page's Gujarati lines into Hindi
• 🌐 हि→ગુ — translate the page's Hindi lines into Gujarati
• 🌐 →EN — translate the page's Gujarati and Hindi lines into English

The translation appears as a block under each translated line (the original stays, so narration and study are unaffected). Each block has its own 🔊 button to read that verse's translation aloud on demand. The translate controls appear only in Text View. Tap the same translate button again to hide. Sanskrit lines are never sent. Results are cached on-device. Uses Google Translate's public endpoint under the same consent as Translate verse. To have translations spoken automatically as part of continuous narration, enable 🔊 Read translation in narration in Settings.

Bookmark notes

Tap the ☆ on any line to bookmark it. In the Library, each bookmark shows the full line text (not just a snippet), and you can attach a note to it (tap "＋ Add note"). Notes and the full text are stored on-device.

Split Sandhi (word separation)

Sanskrit verses fuse words together through sandhi (e.g. तपोवनम् = तपस् + वनम्). The ✂️ Split Sandhi button (on every Sanskrit verse, next to Show Meanings) separates them.

How it works — accuracy first, with honest fallback:

1. Real morphology — it first sends the verse text to the University of Hyderabad Heritage segmenter, an academic Sanskrit morphological analyzer, routed through your own Cloudflare Worker (the same one used for Cloud OCR). The badge shows "University of Hyderabad Heritage segmenter ✓" when this succeeds.
2. On-device fallback — if the service is unreachable, it falls back to a built-in rule-based assistant that handles the reliable cases (avagraha ऽ, visarga sandhi). The badge then says "Guided assistant".

The overlay lets you edit the result directly:

• A live split-out line below the original verse shows the current split joined together, updating as you merge/split
• Each separated word is a chip showing its meaning translated in your chosen language beneath it
• A Meaning in: ગુ Gujarati / हि Hindi / EN English toggle switches all word meanings (defaults to Gujarati; remembers your choice)
• Tap a word to also see its full Monier-Williams dictionary entry, and 🔊 to hear it
• · merge · between any two words joins them back
• ✂ on a word opens an inline picker right in that row — tap a cut-point between letters to split there (no separate popup)
• 🔊 Chant all padas speaks the separated words in sequence
• 💾 Save remembers your split for that verse; 🔄 Re-split from API re-queries; ↺ Reset restores the original

Honest limitation: Sanskrit segmentation is inherently ambiguous — even the academic engine returns the ranked best analysis, not a guaranteed-correct one. And per-word translations come from Google Translate, which is weak on isolated/inflected Sanskrit words — treat the inline translation as a quick gloss and the tap-for-dictionary (Monier-Williams) as the more reliable source. The split/merge editing exists precisely so you have the final say. Treat all of it as a study aid.

Setup: Split Sandhi reuses your Cloud OCR Worker URL. If you haven't set that up (see Cloud OCR setup), it uses the on-device assistant only. Per-word meanings require online lookups to be enabled (same consent as Show Meanings).

Correcting language & text

Automatic script detection is good but not perfect, especially on OCR'd pages. Two manual overrides give you the final say, and both persist (saved in your browser):

Language override — long-press any word (or right-click on desktop) to open a picker:

• Set that word to Sanskrit / Hindi / Gujarati, or back to Auto-detect
• Or set the whole line's language — this is what controls the narration voice (voice is chosen per line)

Use this when a Sanskrit verse is being read in the Hindi voice, or a word shows the wrong dictionary.

Text correction — when OCR garbles a verse, use Re-OCR a region (in Document View) to fix the actual text. Corrections are keyed to the original text and survive re-processing.

Vedabase deep links

For recognized scriptures (Śrīmad-Bhāgavatam, Bhagavad-gītā), each verse gets a 📖 link to the official Vedabase entry:

• Filename hint helps: name files like SB_3.25.pdf or BG_2.pdf for auto-detection
• Source chip shows "SB — set canto/ch" — tap to enter chapter/verse context
• Once context is set, every verse line has 📖 SB 3.25.6 on vedabase.io ↗ pointing to the official commentary

Manual reference entry (when auto-detection fails): if a Sanskrit verse has no working link, it shows a 🔗 Set reference button. Tap it to enter the details by hand:

• Choose the scripture (Śrīmad-Bhāgavatam or Bhagavad-gītā)
• Enter Skand, Adhyay, and Shlok (for BG, just Adhyay and Shlok)
• A live preview shows the resulting vedabase.io link before you save
• Once saved, the verse shows its 📖 link, and a ✎ appears to edit it later

Your manual references are saved (keyed to the verse text) and persist across reloads and re-processing. Use ↺ Clear my reference in the dialog to remove one.

Vedabase content is BBT-copyrighted; we deep-link only, never copy text.

OCR for scanned documents

OCR (Optical Character Recognition) extracts text from images and scanned PDFs.

When to use OCR

Force OCR toggle

Located near "Clear All" on the main screen. Turn ON when:

• You uploaded a scanned PDF and see "No Indic text found"
• The PDF's embedded text is garbage (font mapping broken)
• You want to re-run OCR with a different engine

Local OCR (default)

Uses Tesseract.js running entirely in your browser. Free, private, works offline once language data is cached. Slower (~5–15 seconds per page) and less accurate than cloud OCR on poor scans.

Language data (~5MB for Gujarati + Devanagari) is downloaded from tessdata.projectnaptha.com on first OCR, then cached.

OCR quality indicator

After processing, a small badge near each source shows which engine processed it:

• ☁️ Google (green) — Google Cloud Vision (via your Cloudflare Worker)
• ☁️ OCR.space (green) — OCR.space free tier
• 📱 Local (grey) — Tesseract.js fallback

A confidence banner appears for low-quality results (< 60% confidence) — verify against the original page.

You can tap the engine badge any time to see a per-page diagnostic: which engine OCR'd each page, your current settings, and why cloud failed (if it did).

Re-OCR a region

When OCR mis-reads part of a page — most commonly an inline Sanskrit verse embedded in Gujarati/Hindi prose — you can re-OCR just that area instead of the whole page.

In Document View, each page has a 🔍 Re-OCR a region button. It is pinned to the top of the page (sticky), so it stays visible no matter how tall the page image is — you never have to scroll to the bottom to find it.

1. Tap it, then drag a box over the text that was mis-read
2. A dialog shows the cropped image and runs OCR automatically
3. Choose the language for that region — Sanskrit only (default, best for verses), Gujarati only, Hindi only, or Auto
4. Choose the engine — ☁️ Google Vision (if configured) or 📱 Local
5. Edit the result by hand in the text box if needed
6. Apply it in one of three ways (below)

Three ways to apply the result:

• Replace whole line(s) — tick one or more lines in the checklist and tap ✓ Replace selected line(s). For a multi-line region, tick all the lines it covers; the OCR result is split by line and matched to them top-to-bottom. (More result lines than ticked → the surplus folds into the last line; fewer → the extra ticked lines are removed.)
• Replace only part of a line — tap ✎ Replace only part of a line, choose the line (its full text loads into an editor), select the portion to fix (or place the cursor), tap ⤵ Put OCR result here, then ✓ Save line. Use this when the crop covers just a fragment, so the rest of the line is preserved.
• Single line — tick exactly one line and replace it, as before.

All edits and removals are keyed to the original text and persist across re-processing.

Why this helps: OCR engines assign one dominant script per visual line, so short inline Sanskrit gets misread as the surrounding script. Cropping just the verse and forcing Sanskrit-only removes that conflict and usually reads the verse correctly. If the scan is poor, you can also just type the correction directly — re-OCR is optional.

This works with local OCR too — no cloud account needed for the on-device option.

Fix a single line from Text View

You don't have to go to Document View to fix one line. In Text View each line has a small ✏️ button (next to the ☆); it glows amber when the line contains low-confidence words. Tapping it opens Fix line:

• Edit the text by hand in the box (correct script font), then ✓ Save correction — the fix is keyed to the line's original text and persists across re-processing and reopen, even if you edit the same line again later.
• 🔁 Re-OCR this line — pick an engine (☁️ Google Vision or 📱 Local) and a language (Sanskrit / Gujarati / Hindi / Auto); it crops just that line from the page image, runs OCR, and drops the result into the editable box for review before you save. Recovers a mangled verse without reprocessing the whole PDF.

(Re-OCR needs the line's position on the page image, so it's available for OCR'd PDF/image pages; for pasted or native-text PDFs the dialog offers manual edit only.)

Low-confidence flagging

With ⚠️ Check OCR on (Prefs strip), words the OCR engine reported low confidence for get a wavy amber underline (a tooltip shows the confidence). Both Google Vision and local Tesseract provide per-word confidence; it's stored in the OCR cache so flags persist on reopen.

Cloud OCR setup

Cloud OCR significantly improves accuracy on poor scans, photos, and books with marginalia. Optional and opt-in — disabled by default.

Why a proxy?

Google Cloud Vision blocks direct browser calls (CORS). The app uses a free Cloudflare Worker as a proxy between your browser and Google. Your API key lives on Cloudflare as an encrypted secret — never exposed in browser source code.

One-time setup (≈ 10 minutes)

Two parts:

Part 1: Get a Google Cloud Vision API key

1. Go to console.cloud.google.com
2. Create a new project (e.g. indic-reader)
3. Billing → Link a billing account (free tier covers 1,000 calls/month at $0, but a card is required)
4. Billing → Budgets & alerts → Create budget: set amount to
   with default thresholds — this emails you BEFORE any charge
5. APIs & Services → Library: search "Cloud Vision API", click ENABLE
6. APIs & Services → Credentials → + Create credentials → API key
7. Copy the key. Click "Edit API key":
   • API restrictions: Restrict key → check ONLY "Cloud Vision API"
   • Application restrictions: None (Worker calls don't preserve referrer)
   • SAVE

Part 2: Deploy a Cloudflare Worker

1. Go to dash.cloudflare.com → Workers & Pages
2. Create → Hello World → name it (e.g. indic-ocr-proxy) → Deploy
3. Open Worker → Edit code → delete the template → paste the entire cloudflare-worker-google-ocr.js file
4. Save and Deploy
5. Worker → Settings → Variables and Secrets, add two:
   • GOOGLE_API_KEY = (your key from Part 1) — mark as Encrypt (Secret)
   • ALLOWED_ORIGIN = your app's URL (e.g. https://indic-reader.pages.dev) — use * only for testing
6. Save

Part 3: Configure the app

1. In the app, tap ⚙ Cloud OCR (gear icon near Cloud OCR toggle)
2. Provider: ⭐ Cloudflare Worker → Google Vision
3. Paste your Worker URL (e.g. https://indic-ocr-proxy.YOUR.workers.dev)
4. Tap 🔍 Test connection — should show ✓ Connection works. Detected: "TEST"
5. Save credentials
6. Flip ☁️ Cloud OCR toggle ON

OCR.space alternative (Gujarati only, no setup)

For Gujarati-only documents and quick testing:

1. Get a free key at ocr.space/ocrapi/freekey
2. ⚙ Cloud OCR → choose OCR.space → paste key → save

Free tier: 25,000 pages/month. No Sanskrit support in free tier.

Cost reality

For personal reading (a few dozen pages a day), you'll never exceed any free tier. Set the

This is the user guide for Indic Reader — a free, privacy-first reader, OCR, narrator and translator for Sanskrit, Hindi and Gujarati scripture. Everything runs in your browser.