PDF to MP3 — Convert PDF, EPUB, DOCX & Text to Audiobook

This actor converts PDF, EPUB, DOCX, Markdown, HTML, TXT, and RTF files into MP3 audiobooks using free Microsoft Edge neural TTS (no API key required), with OCR for scanned PDFs, 70+ languages, and optional BYOK OpenAI or ElevenLabs voices. Pricing is pay-per-use: $0.05 to start plus $0.04 per minute of audio produced.

Turn any document into an MP3 audiobook in one run. Paste a URL, upload a file, drop in raw text, or send base64 bytes. Pick a voice. Click run. Get a downloadable MP3. No prompts to chain, no manual chunking, no ffmpeg gymnastics. Just clean audio at the end.

Even scanned / image-only PDFs work — when a page has no text layer, the actor automatically OCRs it (Tesseract) and narrates the recovered text.

Use the free Edge TTS voices by default (400+ neural voices, 70+ languages, no API key) — or bring your own OpenAI / ElevenLabs key for premium voices and steerable narration.

Perfect for: Anyone with a backlog of Project Gutenberg / Calibre EPUBs and zero time to read them, researchers listening to arXiv PDFs on a commute, business users turning Word DOCX reports into audio briefings, devs converting READMEs and blog drafts into audio for proofing, knowledge workers narrating long reports, accessibility-first publishers, podcast producers prototyping audiobook conversions, students reviewing textbook chapters on the go, journalists drafting voiceovers, and Substack writers exporting their newsletter to audio.

Features

Feature	Description
Seven input formats	PDF (text-layer), DOCX (Word, via mammoth), EPUB (ebook, spine-ordered chapters), Markdown (syntax stripped), plain text, HTML (tags stripped), RTF (control codes stripped). Auto-detected from magic bytes / mimetype / extension
OCR for scanned PDFs	Pages with no selectable text layer are auto-rendered and OCR'd with Tesseract (7 languages: EN, ES, FR, DE, IT, PT, NL) so scanned books and photographed pages narrate too. Only the pages that actually need OCR are processed and billed
Encrypted PDF support	Provide the password via pdfPassword to decrypt and narrate password-protected PDFs
SSRF-guarded URL fetch + proxy	Document URLs are validated against private / internal address ranges before fetching. Optional Apify proxy for hosts that block datacenter IPs
Four ways to provide it	Public URL, file upload, base64 paste, or raw text paste (great for blog drafts & ChatGPT replies)
Free Edge TTS by default	Microsoft Edge neural voices, no API key, no per-character cost — works out of the box
One document in, one (or many) MP3s out	Short inputs produce a single MP3. Long ones are auto-split into chapter-sized parts, plus a shareable INDEX.html page with inline players + download links
Multiple TTS engines	Edge TTS (FREE, default), OpenAI gpt-4o-mini-tts (steerable), OpenAI tts-1 / tts-1-hd, ElevenLabs Flash v2.5 / Turbo v2.5
BYOK for premium models	OpenAI and ElevenLabs models require your own API key (we never markup the provider price)
Steerable narration	With gpt-4o-mini-tts you can prompt the voice ("calm audiobook narrator", "energetic podcast host", "slow and deliberate") without retraining
Auto language detection	Edge TTS auto-picks a matching voice based on the text language. Or force a specific Azure voice name (e.g. en-US-AndrewNeural)
Markdown-aware	Markdown syntax is stripped (no "asterisk asterisk bold asterisk asterisk"). Headings, lists, links, code fences, tables, and inline formatting all read as natural prose
Page / section range support	Narrate the whole document, a single chapter, or a custom slice. Works on PDFs (real pages) and TXT / MD / HTML (~3000-char pseudo-pages). Range syntax: 1-10, 1,3,5, 1-3,7-9
Smart chunking	Text is split on paragraph and sentence boundaries before TTS. Hard cuts respect word boundaries so chunks don't start mid-word
Pre-flight cost preview + hard cap	Every run prints an estimated ceiling cost before TTS starts and writes a PREVIEW key. Set maxCostUsd to stop before TTS if the estimate is too high, or to stop narration cleanly at the last part the budget covers if the document runs longer than estimated
Provider-adaptive concurrency	Auto-caps parallel TTS calls per provider (ElevenLabs 2, Edge 8, OpenAI 10) so free tiers don't hit 429 storms
Resume failed runs	Already-synthesized chunks are cached. If a long run times out or fails partway, the next run picks up where it left off — no re-paying for TTS already done
Skip-failed-chunks mode	If a single chunk keeps failing after retries, skip it and keep narrating the rest of the document (configurable). Auth / quota errors always abort cleanly
ffmpeg concat + ID3 tags	Chunks are stitched into valid MP3 containers with correct duration metadata and ID3 tags (title, album, track, genre=Audiobook) so players show proper info. No "audio glitch at minute 4" bugs
Transparent pricing	Pay-per-event: one simple meter — per audio minute produced (plus a small per-run fee, and an OCR surcharge only for scanned pages). No per-page double-billing, no markups on BYOK providers

How to Use

Step 1: Provide a document

Any one of:

• Document URL — any publicly reachable URL ending in .pdf, .docx, .epub, .md, .markdown, .txt, .html, .htm, or .rtf
• Upload file — drag and drop a PDF / DOCX / EPUB / MD / TXT / HTML / RTF (uploaded to the run's key-value store)
• Base64 content — paste raw base64 bytes; format auto-detected from magic bytes / mimetype
• Raw text paste — paste prose, Markdown, or HTML directly into the text field (perfect for blog drafts, ChatGPT replies, READMEs)

Scanned / image-only PDF pages (no text layer) are automatically OCR'd when enableOcr is on (default). For encrypted / password-protected PDFs, supply pdfPassword.

Step 2: Pick a voice and model

Model	Best for	API key required
edge-tts (default)	Free, long books, multi-language	No key — free
openai-gpt-4o-mini-tts	Steerable narration with style instructions	OpenAI key (BYOK)
openai-tts-1	Bulk runs, low cost, supports speed	OpenAI key (BYOK)
openai-tts-1-hd	High-quality OpenAI audio	OpenAI key (BYOK)
elevenlabs-flash-v2_5	Fast, real-time-quality voices	ElevenLabs key (BYOK)
elevenlabs-turbo-v2_5	Highest-quality ElevenLabs voices	ElevenLabs key (BYOK)

Step 3: (Optional) Set a range, voice, speed, or instructions

• voice — leave blank for auto. For Edge TTS use an Azure ShortName like en-US-AndrewNeural, es-ES-ElviraNeural. For OpenAI: alloy, echo, fable, onyx, nova, shimmer, coral, sage. For ElevenLabs: a voice ID.
• language — auto (recommended) or a specific ISO code (Edge TTS only — OpenAI voices are multilingual).
• pageRange — e.g. 1-10 or 1,3,5 or 1-3,7-9. Empty = whole document. For non-PDF formats, "pages" are ~3000-char sections.
• speed — 0.25 to 4.0. Only applies to openai-tts-1 / openai-tts-1-hd.
• instructions — free-form style hint for openai-gpt-4o-mini-tts, e.g. "Calm, slow audiobook narrator with a neutral accent."
• enableOcr — on by default. Auto-OCRs scanned PDF pages that have no text layer. Turn off to fail fast on scans instead.
• pdfPassword — password for encrypted PDFs.
• proxyConfiguration — optional Apify proxy, used only for the Document URL fetch.

Step 4: Run and download

The Actor:

1. Downloads / decodes / reads the input
2. Detects the format (PDF magic bytes + content-type + extension + content sniff)
3. Extracts and normalises the text (page-range aware, Markdown / HTML aware)
4. Splits into TTS-sized chunks at sentence boundaries (word-boundary safe hard cuts)
5. Synthesises each chunk with the chosen provider in parallel
6. Folds chunks into chapter-sized parts as they complete (ffmpeg concat)
7. Uploads each part to the key-value store + writes a shareable INDEX.html

You'll find the result in:

• The dataset — one row with metadata (indexUrl, audioUrl, partsCount, parts[], durationSeconds, chars, pagesProcessed, cost, status)
• The key-value store — each MP3 part, the INDEX.html page, the PREVIEW estimate, and the OUTPUT record

Input Reference

Field	Type	Required	Description
documentUrl	string	one of	Public URL (PDF / TXT / MD / HTML)
documentFile	file	one of	Upload a PDF / TXT / MD / HTML from your device
documentBase64	string	one of	Base64-encoded document bytes
text	string	one of	Paste raw prose, Markdown, or HTML directly
model	enum	no	TTS model (default edge-tts, free)
voice	string	no	Edge ShortName, OpenAI voice, or ElevenLabs voice ID
language	enum	no	Auto-detect (default) or specific ISO code
speed	number	no	0.25 to 4.0, default 1.0. tts-1 / tts-1-hd only
instructions	string	no	Free-form style for gpt-4o-mini-tts
pageRange	string	no	e.g. 1-10 or 1,3,5. Empty = full document
chunkSize	integer	no	500 to 4096, default 4000 (auto-clamped to 2500 for ElevenLabs)
concurrency	integer	no	1 to 20 parallel TTS requests, default 5 (auto-clamped per provider)
resume	boolean	no	Skip already-synthesized chunks from previous runs (default true)
skipFailedChunks	boolean	no	Skip individual chunk failures instead of aborting (default true)
maxPartMb	integer	no	Max size per MP3 part, default 40MB
maxCostUsd	number	no	Hard ceiling (min 0.05). Stops before TTS if the estimate exceeds it, and stops narration at the last part the budget covers if reality overshoots the estimate (status: "capped")
enableOcr	boolean	no	OCR scanned / image-only PDF pages (default true)
pdfPassword	secret	no	Password for encrypted PDFs
proxyConfiguration	object	no	Apify proxy for the Document URL fetch
openaiApiKey	secret	required for openai-*	Your OpenAI API key (BYOK)
elevenlabsApiKey	secret	required for elevenlabs-*	Your ElevenLabs API key (BYOK)
debug	boolean	no	Verbose logs

Output Example

One dataset row per run. It leads with the outcome so a person or an agent can act on it without parsing the rest.

{
    "status": "success",
    "statusMessage": "Narrated the whole document into 2 MP3 parts totalling 30m 43s.",
    "title": "attention-is-all-you-need",
    "indexUrl": "https://api.apify.com/v2/key-value-stores/.../records/INDEX",
    "audioUrl": "https://api.apify.com/v2/key-value-stores/.../records/narration-abc123-part001.mp3",
    "audioUrls": [
        "https://api.apify.com/v2/key-value-stores/.../records/narration-abc123-part001.mp3",
        "https://api.apify.com/v2/key-value-stores/.../records/narration-abc123-part002.mp3"
    ],
    "audioKvKey": "narration-abc123-part001.mp3",
    "durationSeconds": 1843.2,
    "durationHuman": "30m 43s",
    "partsCount": 2,
    "sizeBytes": 19200000,
    "audioFormat": "mp3",
    "parts": [
        { "part": 1, "key": "narration-abc123-part001.mp3", "url": "https://...", "durationSeconds": 1200, "bytes": 12500000 },
        { "part": 2, "key": "narration-abc123-part002.mp3", "url": "https://...", "durationSeconds": 643.2, "bytes": 6700000 }
    ],
    "chars": 48210,
    "pagesProcessed": 24,
    "ocrPagesProcessed": 0,
    "language": "auto",
    "voice": "en-US-AndrewNeural",
    "model": "edge-tts",
    "cost": 1.29,
    "chunksTotal": 13,
    "chunksSucceeded": 13,
    "chunksFailed": 0,
    "generatedAt": "2026-09-07T10:30:00.000Z"
}

status is one of:

Status	Meaning
success	The whole document was narrated.
partial	Some passages failed after retries and were skipped; the gaps are listed in failedChunkIndexes.
capped	Narration stopped early to stay inside a cost ceiling. The audio produced is complete and playable.
failed	No audio was produced. error and errorCode say why. The run still exits successfully.

The MP3 parts also live in the run's key-value store alongside an INDEX page with an inline player and a download link per part, and a PREVIEW record written before any speech is synthesized.

Use Cases

• Any ebook → free audiobook — drop a Project Gutenberg EPUB and listen to a full classic novel. Spine-ordered chapters narrate in the right sequence.
• Word doc → audio briefing — drop your DOCX report and listen on a commute instead of skimming on screen.
• Audiobook prototyping — convert your ebook PDF into MP3 to validate narrator tone before commissioning a human voiceover.
• Research papers on the go — listen to arXiv PDFs during a commute or workout.
• README → audio — paste your project README and listen to your own docs to spot rough explanations.
• Blog draft proofing — paste a Markdown blog draft and listen to it before publishing. You'll hear awkward phrasing you'd never catch reading.
• ChatGPT reply → podcast snippet — copy a long ChatGPT response into the text field and listen as audio.
• Accessibility — generate audio versions of internal documentation for screen-reader-light workflows.
• Onboarding — pipe HR PDFs (handbooks, policies) into audio for distributed teams.
• Newsletter audio versions — automatically narrate weekly reports (PDF, MD, or HTML) for paying subscribers.
• Language learning — narrate text in different voices and speeds to practise listening comprehension.
• Substack → audio export — export a post as HTML and narrate it for a podcast feed.

Use it from an AI agent (MCP)

This Actor is exposed as an MCP tool by the hosted Apify MCP server, so Claude, ChatGPT and any agent framework can narrate a document without writing code. No Standby mode and no special configuration are involved: the server reads the input schema and builds the tool from it.

• Actor ID: marielise.dev/pdf-to-mp3
• MCP tool name: marielise-dot-dev--pdf-to-mp3 (the server rewrites . to -dot- and / to --)
• Server: https://mcp.apify.com — Streamable HTTP only, OAuth or a bearer token. The legacy /sse endpoint has been removed, and there is no anonymous access: every call needs an Apify token, and the run is billed to that account like any other.

Claude Desktop

The simplest route is Settings → Connectors → Add custom connector with the server URL https://mcp.apify.com; your browser opens once to authorise Apify.

To load only this Actor and nothing else, use a tools selector. In claude_desktop_config.json:

{
  "mcpServers": {
    "pdf-to-mp3": {
      "url": "https://mcp.apify.com/?tools=fetch-actor-details,marielise.dev/pdf-to-mp3"
    }
  }
}

If your client cannot do OAuth, authenticate with a token instead:

{
  "mcpServers": {
    "pdf-to-mp3": {
      "url": "https://mcp.apify.com/?tools=fetch-actor-details,marielise.dev/pdf-to-mp3",
      "headers": { "Authorization": "Bearer YOUR_APIFY_TOKEN" }
    }
  }
}

Claude Code

$claude mcp add pdf-to-mp3 "https://mcp.apify.com/?tools=fetch-actor-details,marielise.dev/pdf-to-mp3" -t http

Example tool call

Every input is optional, so an agent can call the tool with almost nothing and still get audio back:

{
  "name": "marielise-dot-dev--pdf-to-mp3",
  "arguments": {
    "documentUrl": "https://arxiv.org/pdf/1706.03762",
    "pageRange": "1-8",
    "voice": "en-US-AndrewNeural",
    "maxCostUsd": 2.0
  }
}

The call returns run metadata and storage IDs rather than the rows themselves. Fetch the result with the auto-injected get-dataset-items tool using the returned datasetId, and you get one row:

{
  "status": "success",
  "statusMessage": "Narrated the whole document into 2 MP3 parts totalling 51m 12s.",
  "title": "1706.03762",
  "indexUrl": "https://api.apify.com/v2/key-value-stores/<storeId>/records/INDEX",
  "audioUrl": "https://api.apify.com/v2/key-value-stores/<storeId>/records/narration-<runId>-part001.mp3",
  "audioUrls": ["...part001.mp3", "...part002.mp3"],
  "durationHuman": "51m 12s",
  "partsCount": 2,
  "audioFormat": "mp3",
  "voice": "en-US-AndrewNeural",
  "cost": 2.09
}

audioUrl and every entry in audioUrls are direct, publicly fetchable MP3 links, so an agent can hand them straight to a player or a download step without pulling a binary through the tool response.

Calling the Actor directly instead

For agents that talk to the Apify API rather than MCP, start a run and take the dataset row back in the same response. Keep the document small when doing this: the synchronous endpoint has a hard 300-second ceiling.

curl -X POST "https://api.apify.com/v2/acts/marielise.dev~pdf-to-mp3/run-sync-get-dataset-items?token=$APIFY_TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"documentUrl": "https://www.orimi.com/pdf-test.pdf", "maxCostUsd": 1.0}'

For a full book, start the run asynchronously and collect the result from a webhook or by polling instead:

curl -X POST "https://api.apify.com/v2/acts/marielise.dev~pdf-to-mp3/runs?token=$APIFY_TOKEN&maxTotalChargeUsd=5" \
  -H 'Content-Type: application/json' \
  -d '{"documentUrl": "https://www.gutenberg.org/ebooks/1342.epub.noimages"}'

Keeping an agent's spend bounded

Narration is billed per minute of audio, and a long book is a lot of minutes. Two independent ceilings exist, and this Actor honours both by stopping rather than by handing over unbilled audio:

• maxTotalChargeUsd on the run — the platform-level limit, and the right one for an autonomous agent to set.
• maxCostUsd in the input — refuses up front if the pre-flight estimate is already over, and stops cleanly mid-run if the document turns out longer than estimated.

Either ceiling produces status: "capped" with the audio generated so far and a statusMessage explaining what happened, never a failed run.

Pricing

This Actor uses Pay Per Event so you only pay for the work the run actually does. No premium-voice surcharges, no provider markups.

Event	Price (USD)	When charged
apify-actor-start	$0.05	Automatically at run start, one event per GB of allocated memory (minimum one) — the per-job base fee, and the only charge on a run that produces no audio
audio-minute-generated	$0.04	Per minute of MP3 actually produced and uploaded — the single value meter
ocr-page-processed	$0.03	Only for scanned / image-only PDF pages that had no text layer and were recovered via OCR. Text-layer PDFs never pay this

Typical cost example — a 20-page research paper (~40k chars, ~45 minutes of audio):

• apify-actor-start: $0.05
• ~45 × audio-minute-generated: $1.80
• Total: ~$1.85 for the full paper (≈ $0.04 per minute of audiobook)

Roughly 900 characters of text becomes one minute of speech, so you can estimate any document before you run it. PRICING.md in the repository shows the measured cost model behind these prices.

For OpenAI / ElevenLabs models, this is all you pay Apify. You pay the provider directly with your own API key on top — that's the whole point of BYOK: no markup.

Pre-flight cost preview & hard cap

Every run writes a PREVIEW key to the key-value store BEFORE TTS starts, with pages to process, estimated audio minutes, and the estimated ceiling cost. The same numbers are printed to the run log.

Set the optional maxCostUsd input to enforce a hard ceiling. If the estimate already exceeds it the run stops before any TTS and you pay only the actor-start fee. If the document turns out longer than estimated (slow speech, CJK scripts, OCR'd text), narration stops at the last part the budget covers and returns status: "capped" with the audio produced so far — the run never hands over minutes it cannot bill for, and never bills past your ceiling. OCR is bounded by the same ceiling before it runs.

Apify's run-level Max total charge works the same way and is the better option for automated and agent callers. Both are honoured; whichever binds first stops the narration.

BYOK — Bring Your Own Key

Model family	Key field	Where to get it	Free tier?
edge-tts (default)	—	No key needed	Yes — completely free
openai-*	openaiApiKey	https://platform.openai.com/api-keys	OpenAI charges per character
elevenlabs-*	elevenlabsApiKey	https://elevenlabs.io/app/settings/api-keys	ElevenLabs free tier available

The actor never logs your keys (isSecret: true) and never proxies your calls through our servers — your key talks directly to the provider from inside the actor's run.

FAQ

Which document formats are supported?

• PDF (.pdf) — text-layer PDFs extract natively; scanned / image-only pages are auto-OCR'd (Tesseract).
• DOCX (.docx) — Word documents, parsed with mammoth. Styles, lists, tables, footnotes handled natively.
• EPUB (.epub) — ebooks. Walked in spine order so chapters narrate in the right sequence. HTML stripped per chapter.
• Markdown (.md, .markdown, .mdx) — syntax stripped so the voice reads natural prose.
• Plain text (.txt, .text) — UTF-8, BOM handled.
• HTML (.html, .htm, .xhtml) — tags stripped, entities decoded, scripts and styles removed.
• RTF (.rtf) — control codes stripped, unicode escapes and hex bytes decoded.

Does it work on scanned PDFs?

Yes. When a PDF page has no selectable text layer, the actor renders it (poppler pdftoppm) and runs OCR (Tesseract) to recover the text, then narrates it. OCR runs only on pages that need it, and those pages are billed via the ocr-page-processed event ($0.03/page). Built-in OCR languages: English, Spanish, French, German, Italian, Portuguese, Dutch (others fall back to English). Turn it off with enableOcr: false to fail fast on scans instead.

Does it work on password-protected PDFs?

Yes — pass the password in the pdfPassword input and the actor decrypts the PDF before extraction.

What about ODT, MOBI, AZW3, or Pages?

Not supported in v0.1. Convert ODT to DOCX first; for Kindle formats, convert via Calibre to EPUB.

Why do OpenAI / ElevenLabs models require BYOK?

So we never markup the provider price. Pay Apify for the actor work, pay OpenAI / ElevenLabs directly for the TTS calls. Cleaner, cheaper, more honest. For zero-key zero-friction runs, the default edge-tts is free and gives great quality on 70+ languages.

How is "page" defined for TXT / MD / HTML?

There are no real pages, so the actor splits the cleaned text into ~3000-char pseudo-pages — roughly the length of one PDF page of prose. This is only used for the pageRange selector; billing is per audio minute, so it's format-agnostic and fair either way.

How long can the document be?

PDFs up to 50 MB. EPUB up to 40 MB. DOCX up to 30 MB. TXT / MD / HTML / RTF up to 20 MB of decoded text. There is no hard page limit. Long inputs are auto-split into chapter-sized MP3 parts (configurable via maxPartMb).

What if my run times out or fails partway?

Re-run with the same input. The resume option (on by default) skips already-synthesized chunks via a shared cache, so you only pay TTS for the missing pieces.

Can I get word-level timestamps?

v0.1 does not emit timestamps. Coming in a future version.

Can I use multiple speakers / podcast mode?

Not in v0.1. Single-voice narration only.

What audio format is produced?

MP3, standard playback on any device. Edge TTS produces 24 kHz mono; OpenAI / ElevenLabs use their default high-quality output.

Can I override the voice with a custom ElevenLabs voice?

Yes — paste any ElevenLabs voice ID into the voice field when using an elevenlabs-* model. You can clone your own voice in your ElevenLabs account and use that ID here.

Is my API key safe?

Yes. API keys are marked isSecret: true in the input schema and are never logged or persisted.

Built and maintained by Equipinico

Need a custom variant (different language model, custom voices, SSML support, podcast multi-speaker, EPUB / DOCX support)? Reach out via the Apify Store contact link.