Show HN: I generated 70k audiobooks with OpenAI Text-to-Speech

Original Article ↗ Hacker News Discussion ↗

An indie developer has built a service that can turn around 70,000 Project Gutenberg titles into on‑demand audiobooks using OpenAI’s text‑to‑speech, prompting debate over technical quality, user experience, and where AI narration already works well (notably non‑fiction) versus where it still falls short (fiction and acting). Commenters probe the business model and ethics of charging for AI renditions of public domain works, suggesting alternatives like crowdfunding, pay‑once‑unlock‑for‑all, or donation funding, while also questioning a promotional claim of having “generated 70k audiobooks” when audio is created just‑in‑time. Others compare rival TTS systems, request features such as multiple voices and better search, and note that open‑sourcing or using cheaper/open models could change the economics of large‑scale audiobook generation.

Implementation & Open Sourcing

  • Code is currently closed source; some ask for open-sourcing to self-host or contribute.
  • Author describes it as a straightforward wrapper around OpenAI TTS + Google OAuth + a payment provider.
  • Others note there’s limited upside to open-sourcing unless for portfolio/visibility.

User Experience & Feature Requests

  • Requests: search across the large catalog, more login options beyond Google, light theme, and mobile apps.
  • Multiple voices and voice selection (by author or protagonist gender, or per-character) are highly requested but deprioritized so far.
  • Other ideas: multi-narrator “audio play” style, 1.5x generation speed (not just playback), Apple Pay support.

Audio Quality, Languages & Use Cases

  • Several listeners praise the natural cadence; some say it’s the best TTS they’ve heard, especially for essays and non-fiction.
  • Others find it still slightly unnatural, with odd pauses/emphasis, and say it fails badly on poetry/dramatic works (e.g., Shakespeare’s meter).
  • Consensus in the thread: TTS is currently much better for history/philosophy/science/non-fiction than for fiction and dialogue-heavy texts.
  • OpenAI TTS is reported as weak for non‑English; some note other models do better at emotion but worse in voice quality or hallucinations.

Generation Strategy & Scalability

  • System splits books into ~4k-character chunks due to API limits, generating audio on-demand.
  • It pre-generates the next chunk near the end of the current one to keep playback seamless.
  • Full-book pre-generation and chapter MP3s are planned but not finished.

Monetization, Pricing & Caching

  • Current model: one-time purchase of listening “hours,” with pricing set around 50% of raw API costs; profit appears only after multiple purchases of the same book.
  • Revenue so far is very low; author hopes to reach modest MRR.
  • Ideas from commenters:
    • Monthly subscription and mobile app for recurring revenue.
    • Crowd-funding per book (many small contributors unlock a free public audio).
    • First buyer funds generation; others pay less, or listen free.
    • Using the project as a free/donation-based “lead magnet” for other products.

Ethics & Value of Charging for Public Domain

  • Some see charging for public-domain audiobooks as unethical or “gross”; others reply that API/storage costs must be covered and point out that many businesses charge for public-domain content.
  • A compromise suggested: charge only at cost, or let users “donate” generated audio to the public.

Comparisons to Existing Projects & Tools

  • Mention of Microsoft’s prior Gutenberg TTS effort; some say its voices are worse than OpenAI’s.
  • Librivox is cited as a human-read alternative; some prefer human narration, others find many Librivox readings lower quality than the AI.
  • Various TTS engines are discussed (ElevenLabs, Piper, Bark, xTTS, Voicebox); consensus is that OpenAI TTS is currently among the most pleasant but not perfect.

Marketing & Title Controversy

  • A subthread argues over the HN post title claiming “generated 70k audiobooks” since books are generated on demand, not precomputed.
  • Critics call this misleading or a “lie”; supporters say it’s a reasonable shorthand since all 70k are playable via the system and the on-demand detail is disclosed in the post.