Kenkui

Kenkui turns EPUB books into M4B audiobooks. Speech is synthesized locally on your CPU with Pocket-TTS, and a book can be read by a single narrator or by a full cast, with a different voice for each character.

import kenkui as kk

if __name__ == "__main__":
    kk.load_voice("eponine")
    kk.magic_run("book.epub", narrator="eponine")  # writes book.m4b

• One narrator or a full cast. Characters are discovered locally with spaCy or by a language model, dialogue is attributed to its speaker, and voices are cast deterministically, staying consistent across a whole series.
• Fix it without re-rendering. Read the script a book will be spoken from, correct a speaker, a pause, or a pronunciation, and preview only that line.
• Local and predictable. Speech synthesis never touches the network. Voices are downloaded only when you ask for them, and a failed or cancelled render never leaves a half-written file behind.
• A typed, immutable Python API. Every step returns a new pipeline you can branch, inspect, and reuse.

The Kenkui suite

Install

Kenkui needs Python 3.11–3.13, and FFmpeg 5 or newer on your PATH.

pip install kenkui

brew install ffmpeg                  # macOS
sudo apt-get install -y ffmpeg       # Debian / Ubuntu

The install is large, because Pocket-TTS depends on PyTorch. Voices are downloaded separately, on first use: about 225 MB per language, plus about 6.5 MB per voice. To discover characters offline, also install the spaCy extra:

pip install "kenkui[spacy]"
python -m spacy download en_core_web_lg

Usage

One narrator

import kenkui as kk

if __name__ == "__main__":
    kk.load_voice("eponine")  # one-time download; later runs are offline

    book = kk.book("book.epub")
    for chapter in book.inspect().chapters:
        print(chapter.id, chapter.title)

    result = (
        book.pronounce(numbers="standard")
        .pauses(chapter_ms=1500, paragraph_ms=300)
        .assign_voice("eponine")
        .tts()
        .metadata(title="My Book", author="An Author")
        .write("book.m4b", on_event=print)
    )
    print(result.output, result.stats.duration_ms)

Rendering starts worker processes, and each one re-imports your script, so keep rendering under if __name__ == "__main__":.

A full cast

Quote attribution uses any LiteLLM model. Set the provider's API key in your environment, for example OPENROUTER_API_KEY.

import kenkui as kk

MODEL = "openrouter/deepseek/deepseek-v4-flash"

if __name__ == "__main__":
    for voice in ("eponine", "paul", "anna", "charles", "jane", "george"):
        kk.load_voice(voice)

    (
        kk.book("book.epub")
        .series("my-series", book=1)  # characters keep their voices in later books
        .infer_characters("spacy")  # or a LiteLLM model
        .attribute_quotes(MODEL)
        .assign_voices(narrator="eponine", unknown="paul", cast={"alice": "anna"})
        .tts()
        .write("book.m4b")
    )

Correct one line and hear the fix

book = kk.book("book.epub").annotations().assign_voice("eponine")
where = {"chapter": "*", "paragraph": 1}

for row in book.script().at(where):  # what will be said, and why
    print(row.path, row.character, row.provenance, row.text[:60])

book = book.attribute("irulan", where=where).silence(900, where=where)
book.select(where).preview("probe.wav")  # renders only those lines
book.write_annotations()  # saved beside the EPUB for the next render

Examples

examples/ contains runnable scripts, from inspecting a book to batch-rendering a library. Most take the path to an EPUB as their only argument:

Documentation

Full documentation is at https://d1zzl3d0p.github.io/kenkui/:

• Getting started
• Guide: every feature, in depth
• API reference
• Models and voices: the catalog, and voice rights
• Troubleshooting: stable error codes

Contributing

See CONTRIBUTING.md. Changes are recorded in CHANGELOG.md.

License

Kenkui is licensed under the Apache License 2.0. Voices and model weights are separate works with their own terms, and every built-in voice ships marked as not cleared for commercial use. See Models and voices.