The workflow extracts a book, detects its chapters, generates speech in different voices and languages, and prepares it for playback and download.
Role
Product, design, and engineering
Period
Current
Status
Live
Disciplines
Product, Audio systems, AI systems, Engineering
A reader begins with an EPUB or PDF rather than conversion settings.
The source is analyzed before a voice or processing tier is selected.
Voice, language, quality, and estimated cost stay visible before conversion.
A final review makes the selected voice, output, and price explicit before payment.
Long-running synthesis reports progress while queued work completes.
The result is playable, downloadable, and organized around detected chapters.
Detected chapters become direct navigation points in the finished audio.
Book chat answers from the source text alongside the listening experience.
01 · Context
What prompted it
Turning a long document into useful audio is more than text-to-speech: book structure, processing reliability, chapter timing, playback, and temporary file storage all have to work together.
EPUB and PDF sources vary widely, while long-running conversions must survive partial failures and still produce an audiobook that is easy to navigate.
02 · Approach
How it works
Extract and normalize source text, detect chapter structure, then divide the work into processing jobs that can run in batches.
Generate audio through standard and premium pipelines, assemble the result, and audit chapter timing before playback.
Treat the player and chapter model as part of the conversion system, with book chat and privacy-focused retention built around the finished audiobook.
03 · Decisions
Choices along the way
01
Shipped
Process long books as queued jobs
Constraint
A full book exceeds the practical duration and payload of a single serverless request.
Choice
Split normalized text into batches and use a supervised worker queue for parallel processing and retries.
Result
Conversions can progress incrementally and recover at the job level instead of restarting the entire book.
02
Shipped
Make chapters part of the audio model
Constraint
A single audio file without reliable positions is difficult to navigate as a book.
Choice
Detect chapters, retain their metadata, and refine start times through SSML timepoints and audio auditing.
Result
The built-in player can expose chapter navigation alongside ordinary playback controls.
03
Shipped
Delete conversion files automatically
Constraint
Uploaded books and generated audio should not remain in storage indefinitely.
Choice
Apply a documented 30-day retention policy to source files, audiobooks, and conversion metadata.
Result
The live system has a defined, automated lifecycle for conversion data.
04 · Current version
Where it is now
A live EPUB and PDF conversion flow with automatic chapter detection and multiple languages and voices.
Standard and premium processing paths feed a built-in audiobook player with chapter navigation.
Completed books support playback, download, and AI chat about book content.
