Lorenzo, founder of Cirrela, recently detailed the engineering challenges behind personalizing text-to-speech (TTS) voices for children’s names in holiday videos. While the product generates a vertical video of Santa or an elf speaking a child’s name for €6.99, the core technical hurdle isn’t the video generation itself—it’s ensuring the AI pronounces unique names like "Dada," "Siobhán," or "Nguyễn" correctly. Lorenzo notes that standard TTS models, trained primarily on common dictionary words, fail catastrophically on proper nouns, leading to user dissatisfaction and refund requests.

The Failure of Standard TTS Pipelines

The root cause of mispronunciation lies in how TTS models process input. Because names lack dictionary entries, models rely on phonetic guessing, which often results in errors such as "Dada" becoming "Dah-dah" or "Holly" turning into "Hall-ee." Lorenzo highlights that these issues are exacerbated by lip-sync models, where incorrect audio forces the avatar’s mouth movements to match the wrong sounds, making post-generation audio fixes impossible. A recent Reddit thread cited by the founder illustrated this, where a parent rejected a competitor’s output because the voice butchered the name "Dada" and the video length was incorrect, ultimately settling for a "close enough" $5.99 alternative.

Bounding Retakes with Edit Distance Logic

To mitigate these issues without exploding infrastructure costs, Cirrela implemented a three-layer system centered on controlling the TTS input. First, the script uses a {name} slot filled at render time, allowing for precise phonetic respelling (e.g., typing "Siobhan" as "shiv-AWN"). Second, they offer one free re-render if the pronunciation is wrong, but this fix is strictly bounded. The system calculates the edit distance between the original and corrected spelling; changes that alter the name’s identity (like "Léa" to "Lou") are rejected, while minor phonetic adjustments (like "Mael" to "Maël") are allowed. This prevents users from farming free renders by changing names entirely and ensures the video remains relevant to the specific child.

Infrastructure and Future Optimizations

The technical stack relies on HeyGen’s avatar pipeline running on fal.ai, with localized voice presets across eight languages including English, French, Spanish, and Hindi. Lorenzo admits that their current approach—offering a full free preview with a watermark—is expensive, costing a full render per curious visitor. He identifies a clear optimization for the future: implementing a five-second audio-only pronunciation preview. This would allow parents to confirm the name’s sound for pennies before committing to the full video render, a change inspired directly by the user complaints highlighted in the Reddit discussion.

Key Takeaways

  • Demo videos using common names like "Emma" hide TTS failures; test with edge-case names like "Dada" or "Aiko."
  • Implementing edit distance checks for name corrections prevents abuse of free retake policies.
  • Phonetic respelling inputs allow users to guide TTS models without requiring technical knowledge.
  • Audio-only previews significantly reduce rendering costs compared to full video previews.

The Bottom Line

Pronunciation is a critical feature for any AI product that speaks user data, not just a cosmetic detail. Builders must prioritize phonetic control and bounded correction workflows to maintain trust and reduce infrastructure waste.