Your book finds its voice.

The text of each take

New generations keep their text and preparation reference. An association after listening keeps a user declaration for the current text and edit; it does not modify the audio and does not prove the words. The section still needs to be approved afterwards. For old takes, the marker text can serve as a limited comparison.

Book production plan

Local forecast, nothing is sent to the provider. Adjust your assumptions, then recalculate. Audio durations must be measured for this edit; a change invalidates them.

Space and left/right arrows when the player has focus, outside fields and buttons. The export contains the WAV of the rough edit and the text linked to this take; nothing is put online. Limit: 180 MB of WAV per section. Unzip the ZIP to open index.html. Partly cut passages must be fixed before exporting.

Prepare what the voice should read.

Example: “Dr.” → “Doctor”. The manuscript is kept; only a copy prepared for synthesis uses these readings. Character tags are processed before this dictionary.

These are plain-text aliases, not a phonetic alphabet or SSML. The final pronunciation depends on the voice: listen to a trial. Review-lexicon entries are not enabled automatically.

100 rules maximum. The longest terms take priority; replacements are not re-applied in cascade. Active rules apply to every section of the book. A change may make saved fragments incompatible and require regenerating them.

Find a sentence by ear.

Automatic markers follow the synthesis fragments, including pauses between passages. They are neither aligned words nor a transcript of the audio. Old takes have no automatic markers; you can add some by hand. Proofread and listen to each passage before distributing.

Fully removed passages are left out of the export. A partly cut passage blocks the export: correct its text and times, or delete the marker. The audio is unchanged.

Markers relative to this section’s edit. Exported separately from the audio tracks; they are not added to the submission ZIPs. MP3/AAC encoding may introduce a slight offset depending on the player. Long texts are not automatically split into short subtitles.

Who tells each passage?

Create the book’s roles, choose their voices, then insert their names in square brackets, alone on a line. The tag applies until the next role and will not be spoken. A start without a tag uses the first role in the list.

This preparation stays local. Loading voices needs a connection to the voice service; generating then asks for your usual authorisation. The book’s roles are shared by all sections. An undefined role voice inherits the section’s voice, then the book’s.

Staging is manual: no automatic speaker detection. Proofread the text and the voices. Old takes remain available but will need to be compared with the new script. Limits: 30 roles per book, 20,000 characters and 300 passages/fragments per section.

The best of each workshop.

Comparison of the features announced by the publishers, consulted on 18 October 2026. This studio is an independent application, designed to complement the Audi-book workflow.

What works here

Import of TXT/Markdown, Word DOCX, EPUB and text or scanned PDFs with optional OCR and correction before import, pronunciation lexicon, active reading dictionary with preview and TXT/CSV/JSON scripts, editable text/audio passages and SRT/WebVTT exports, exportable review notebook, assisted detection of pauses and peaks, selection loop, listening speed, recovery of microphone segments, management of several books and storage space, microphone choice and guided test, AI voice conversion after server configuration, chapters, teleprompter, microphone, import of audio your browser can decode, reversible trimming and cuts, undo, markers, pickup with pre-roll, microphone meter, profile-based spectral noise reduction, mono LUFS and 4× estimated true peak, local processing, RMS/peak measurements, multiple takes, saving and ZIP in MP3 CBR or WAV. A synthetic sound test lets you check the audio chain.

Limits of this version

The M4B export depends on the browser’s AAC encoder. Text/audio markers by generated fragments or manual entry, without audio transcription or word-by-word alignment. French/English OCR is optional and requires an initial download. AI generation requires a key and engine access configured on the server; costs and rights depend on the provider account. Spectral noise reduction suits steady noise, not conversations or reverberation. Mono LUFS and the 4× estimated true peak are not a certified instrument. Long books must be worked on chapter by chapter; decoding and exports use the device’s memory. The browser’s save can be erased: download your project regularly.

Audio measurements

Mono LUFS with K-weighting and double gate, true peak estimated with the 4× FIR filter of annex 2 of ITU-R BS.1770. These measurements are not a certification. The filter formulas at different frequencies were checked against libebur128; MIT notice.

Reading manuscripts

Local extraction of text PDFs with PDF.js 6.4.299, distributed unmodified. Apache-2.0 licence ; notices for the CMaps and fonts in the vendor/pdfjs folder. Password-protected documents are not supported. Local French/English OCR with Tesseract.js 7.0.0, engine and language models served by KNPLIV itself, loaded with your permission; the pages stay on this device. Apache-2.0 licence and notices in vendor/tesseract.

Local extraction of DOCX and EPUB text. Decompression with fflate 0.8.3, unmodified; MIT licence. Encrypted EPUB resources and old .doc files are not supported.

MP3 encoder

lamejs 1.2.1 / LAME, loaded locally unmodified, under LGPL-3.0. LAME project · LGPL licence · GPL licence · Notice · Encoder sources. The vendor/lame.min.js file can be replaced by a compatible version.

A space for each book.

Each book keeps its texts, takes, edits and settings. The current book is saved before a switch. Importing a project creates a new local copy.

Passages to listen to again.

These indications do not replace listening and look for neither reading errors nor all clicks or noises. Choose the suggestions to keep as markers of the active take. The text and audio are unchanged.

Check the replacements.

The text will be modified only in the chosen sections. Audio takes do not change; modified sections will go back to “to listen”. An existing voice draft will need to be regenerated if its text no longer matches.

Prepare the text to narrate.

Selectable text or scanned pages with optional OCR. Column order, notes and special characters must be checked after extraction.

Excluded areas may contain useful text: set them carefully. Extraction in batches of 500 pages, document of 2,000 pages and 30 MB maximum.

Choose the sections to narrate.

Expand a section to correct its title and text before adding it to the project. Headings stay in the text to be read. DOCX / EPUB / PDF: 30 MB maximum; extracted text: 5 million characters.

Find the saved segments.

Segments received from the microphone are kept on this device during the take. After an interruption, playing them depends on the browser’s format: a partial file may need repair. Only segments already saved can be recovered.

“Add to project” creates a new section. A pickup fragment still needs realigning. Download the original if the browser cannot decode it.