How Abogen Generates 3,000‑Character Audio in 11 seconds Offline – 4.8k‑Star GitHub TTS Tool

Problem statement

Four common drawbacks of existing TTS services: per‑character cloud charges, privacy risk from uploading text, robotic‑sounding output with poor punctuation handling, and lack of support for formats beyond plain TXT.

Project overview

Abogen (Audiobook Generator) is an MIT‑licensed Python project providing a PyQt desktop GUI and a Flask‑based web interface. It uses the lightweight Kokoro‑82M TTS model to generate natural‑sounding speech on modest hardware.

Performance measurements

On a laptop with an RTX 2060, 3 000 characters are synthesized in 11 seconds, producing a 3 min 28 s audio file.

A 50‑page English PDF is fully processed in 5 minutes, yielding a 21‑minute audio file with chapter markers.

CPU‑only environments run slower but retain full functionality and offline capability.

Privacy‑first design

Model and voice packages can be pre‑downloaded. Network requests to Kokoro or HuggingFace can be disabled, ensuring that no text or audio leaves the local machine.

Core functionality

Drag‑and‑drop of EPUB, PDF, TXT, Markdown, SRT, ASS, VTT; automatic chapter detection for EPUB/PDF and metadata injection (title, author, cover) into M4B files.

Custom markers <<CHAPTER_MARKER:章节名>> and <<METADATA_XXX>> for manual segmentation and metadata definition.

Timestamp scripts containing HH:MM:SS are parsed to generate aligned narration for short‑video scripts.

Subtitle synchronization: word‑level (English only) with millisecond precision; sentence‑level for other languages. Export formats include SRT and multiple ASS styles.

Voice mixer blends native male/female voices (e.g., 70 % male + 30 % female) and saves configurations.

Supports nine languages (US/UK English, Mandarin, Japanese, French, Spanish, Italian, Brazilian Portuguese, Hindi). Japanese and Chinese require the optional misaki dependency.

Batch queue processing with independent speed, voice, subtitle style per file; global overrides; real‑time progress logs.

Audio output options: WAV, FLAC, MP3, high‑compression OPUS, M4B (with chapter markers for Audiobookshelf).

Installation methods

Windows one‑click installer (bundles Python, CUDA, dependencies).

Cross‑platform installation with uv (recommended):

uv tool install --python 3.12 abogen[cuda] \
    --extra-index-url https://download.pytorch.org/whl/cu128 \
    --index-strategy unsafe-best-match

Docker container:

docker build -t abogen .

docker run --rm -p 8808:8808 -v ~/abogen-data:/data abogen

Source installation in editable mode:

uv pip install -e .

Typical usage workflow

Drag or paste an EPUB/PDF/TXT into the input area.

Configure language, voice mix, speed, subtitle granularity, output format, and destination folder.

Press Start; conversion runs locally and the output folder opens with audio and subtitle files.

Known limitations

Sentence breaks for English abbreviations (e.g., “Mr.”, “Dr.”) can be reduced by enabling spaCy segmentation.

Long passages may lack emotional variation.

Punctuation‑driven pauses (ellipsis, multiple commas) are still being refined.

Word‑level subtitles are available only for English; other languages are limited to sentence‑level.

Web‑only features (LLM text normalization, Audiobookshelf integration) are not yet present in the desktop GUI.

Windows does not support AMD GPU acceleration; AMD users must use Linux ROCm.

Typical deployment scenarios

Personal e‑book listening: desktop GUI for one‑click EPUB import and M4B export.

Short‑video AI dubbing: web UI with LLM‑driven text cleanup, English word‑level subtitles, and OPUS output.

Audiobookshelf library management: Docker‑deployed web service that uploads finished audiobooks directly to the library.

Secure corporate environments: pre‑download models, disable network, run batch conversions on air‑gapped machines.

Developer extensions: install via uv in editable mode to customize the TTS pipeline.

Repository

https://github.com/denizsafak/abogen