Audiobook Generator

A browser-based audiobook generator that converts EPUB, PDF, HTML, and TXT files into audio using AI-powered text-to-speech — running 100% in your browser, no server required.

Note: Works with DRM-free files only. Files purchased from Apple Books, Google Play Books, or Amazon Kindle with DRM enabled are not supported.

Features

TTS Models

• Kokoro-82M — 27 high-quality English voices, runs via ONNX (WASM or WebGPU)
• Piper TTS — multilingual support, auto-selects voice based on detected language
• Web Speech API — browser built-in, no model download required

Reading & Playback

• Text Reader — read and listen simultaneously with sentence-level highlighting
• Progressive playback — audio starts playing as segments are generated, no waiting for the full chapter
• Click-to-play — click any sentence to start generation and playback from that point
• Resume progress — saves your position per book across sessions
• Keyboard shortcuts — Space, arrow keys, F for fullscreen, ? for help
• Themes — light, dark, and sepia reader themes
• Playback speed — 0.75×–2.0×

Export

• MP3 — universal compatibility, recommended
• M4B — audiobook format with chapter markers
• WAV — lossless, for archival or further processing
• EPUB3 — with Media Overlays (synchronized text highlighting in compatible readers)

Library

• Books saved automatically to IndexedDB, persist across sessions
• Search, sort, and manage your library
• Storage usage indicator

Input

• Upload EPUB, PDF, HTML, or TXT files
• Paste a URL to fetch and convert any article

Quick Start

Prerequisites: Node.js 20+, pnpm

git clone https://github.com/Cabeda/audiobook-generator.git
cd audiobook-generator
pnpm install
pnpm dev

Open http://localhost:5173.

Usage

1. Import — upload a file or paste a URL
2. Select chapters — choose which to convert
3. Pick a model and voice — Kokoro, Piper, or Web Speech
4. Generate — click a chapter to start; audio plays as it generates
5. Export — download as MP3, M4B, WAV, or EPUB3 when done

Development

pnpm dev           # dev server
pnpm build         # production build
pnpm preview       # preview build

pnpm test          # unit tests (vitest)
pnpm test:e2e      # E2E tests (Playwright)
pnpm lint          # ESLint + Prettier
pnpm type-check    # TypeScript check

Project Structure

src/
├── components/
│   ├── TextReader.svelte        # Reader with sentence highlighting & playback
│   ├── AudioPlayerBar.svelte    # Persistent playback controls
│   ├── BookView.svelte          # Chapter list and generation controls
│   ├── LibraryView.svelte       # Library management
│   ├── LandingPage.svelte       # Entry point
│   ├── SettingsPage.svelte      # App-wide TTS settings
│   └── UnifiedInput.svelte      # File/URL input
├── lib/
│   ├── audioPlaybackService.svelte.ts  # Playback engine (Svelte 5 runes)
│   ├── audioConcat.ts                  # WAV/MP3/M4B concatenation
│   ├── services/
│   │   └── generationService.ts        # TTS orchestration, segment batching
│   ├── kokoro/                         # Kokoro TTS client
│   ├── piper/                          # Piper TTS client
│   ├── parsers/                        # EPUB, PDF, HTML, TXT parsers
│   ├── epub/                           # EPUB3 + Media Overlay export
│   └── utils/                          # Voice selection, language detection
├── stores/
│   ├── segmentProgressStore.ts  # Per-segment generation state
│   ├── audioPlayerStore.ts      # Playback state
│   ├── bookStore.ts             # Book and chapter state
│   └── ttsStore.ts              # TTS settings
e2e/                             # Playwright E2E tests

Architecture

Stack: Svelte 5 (runes) + TypeScript, Vite, ONNX Runtime Web, Web Audio API, IndexedDB

Generation flow:

Content (EPUB/PDF/HTML/TXT)
  → parse into chapters
  → segment HTML into sentences (DOM-based, preserves inline markup)
  → generate audio per segment (Kokoro / Piper / Web Speech)
  → stream segments to playback as they complete
  → batch-save to IndexedDB
  → concatenate into final audio file on export

Playback modes:

• Merged audio — single file with time-based segment tracking (smooth seeking)
• Progressive — per-segment blobs chained as generation completes
• On-demand — generate segment on click, buffer ahead

Performance

Mode	Generation speed
Kokoro WASM	~0.5–1.0s per sentence
Kokoro WebGPU	~0.2–0.5s per sentence
Piper	~0.3–0.8s per sentence
Web Speech	instant (browser native)

Model first load: ~5–10s (downloads ~82MB, cached in IndexedDB after).

Troubleshooting

• Out of memory — use q8 quantization instead of fp32, generate fewer chapters at once
• Slow generation — enable WebGPU in Chrome 113+, close other tabs
• Wrong language voice — Piper auto-selects by detected language; override in Settings
• MP3 encoding fails — try WAV format; check browser console for FFmpeg errors

Roadmap

• Sentence-level highlighting in Text Reader
• Progressive playback (listen while generating)
• Piper TTS multilingual support
• EPUB3 Media Overlay export
• Local library with persistent storage
• Resume reading progress
• Adaptive quality (mobile vs desktop)
• Export/import library backup
• Batch processing multiple files

License

MIT — see LICENSE.

Third-party licenses: Kokoro-82M (Apache 2.0), Piper (MIT), lamejs (LGPL), Svelte (MIT), ONNX Runtime (MIT).

Acknowledgments

• Kokoro-82M by hexgrad
• kokoro-js
• Piper TTS
• espeak-ng
• lamejs

---

Author: José Cabeda · GitHub Issues