What needed to change.
Most text-to-speech pipelines read books literally. They do not reason about characters, narrator consistency, dialogue, emotion, accents, or how a long-form story should sound as an audiobook.
Generative AISelected work
AI Audiobook Studio
PDF-to-audiobook system that interprets characters, casts voices, generates expressive narration, and assembles multilingual audio.
What I built or explored.
Built a pipeline that converts PDFs into structured text, uses an LLM to interpret characters and narration, assigns voices, generates speech through TTS providers, and assembles audio chapter-by-chapter. The system also explores multilingual output, pronunciation guidance, conversational delivery, and background music.
Why it matters.
Developed an end-to-end generative-media workflow and iterated through real production issues including long-context failures, invalid structured outputs, voice consistency, API timeouts, language handling, and Docker deployment.
TOOLS / THEMES
FastAPIDockerLLMsSarvam AITTSPython