Lullaby and white-noise apps offer a catalogue of songs in English, Spanish and a few other big languages, and personalised-song gift sites put a baby's name into a stock English tune. Neither helps a diaspora parent in Toronto sing to their baby in Yoruba or Tagalog, the language the grandparents speak and the parent half-remembers. Heritage Lullaby writes an original song with the baby's name in that language, has it checked by a native speaker, sings it, and teaches the parent to sing it themselves, line by line.
The data it runs on
Lyrics come from a frontier LLM prompted with a small per-language style guide (imagery, address forms, tone rules) written by our reviewers. Every song is checked by a paid native-speaker reviewer before release: grammar, tone marks, and "would a grandmother actually say this". The singing voice comes from a singing-voice synthesis model; pronunciation help uses TTS where a good voice exists. Limits: LLM quality drops sharply for lower-resource languages (Yoruba tone marks are often wrong, Igbo and Gujarati are weaker still), and many TTS engines have no Yoruba or Igbo voice at all. Fallback: the reviewer records a plain spoken guide track, and languages without a reviewer stay on a waitlist instead of shipping unchecked text.
1. The mission
No child should grow up never hearing the language their grandparents dream in. We give every parent a lullaby in that language, with their baby's name in it, and the confidence to sing it every night.
2. Why people can't resist it
Guilt turned into a ritual. Many diaspora parents feel they are losing the language. Singing one song nightly is a small, doable act of keeping it.
The name moment. Hearing your baby's name sung in Grandma's language makes parents cry. They replay it and send it.
Grandparents in the loop. Grandparents receive the song and sing back. The family chat lights up every week.
3. How it works for real
Brief: name, language, mood, length, who sings (parent, grandparent), any words to include.
Write: a frontier LLM (Claude or GPT via OpenRouter) drafts original lyrics in short, repeatable lines with an English translation, following the language style guide. Rule: never reproduce traditional lullaby lyrics.
Check: a second model pass (Gemini Flash or DeepSeek) flags grammar, diacritic and tone-mark issues, then a human native speaker approves or edits in a reviewer queue. Paid per song.
Melody: a small rules engine builds a gentle melody in a narrow range (easy to sing tired) at 60 to 70 BPM, mapping Yoruba tones to pitch contour so the melody does not fight the words.
Sing: a singing-voice synthesis model renders the vocal (open models such as DiffSinger-style systems on a GPU, or a hosted music model); TTS voices read lines slowly for practice.
Teach: line-by-line respelling, tone marks and slow playback. Parent records a take; a speech model scores timing, not talent.
4. Cost per use
About $0.50 per song: lyrics and checks about $0.03, singing voice render about $0.15, native-speaker review about $0.30 when spread across reused verses. Practice TTS is fractions of a cent.
Free: one song with a spoken guide (no sung vocal) and the pronunciation guide. Paid ($4.99 a month or $29 a year): sung versions, new verses as the baby grows, grandparent sharing, family voice library.
5. Growth loop
Every song is made to be sent to grandparents. Grandparents forward it to aunts and cousins with babies of their own. Each share link plays the song and offers "Make one for your little one" in the same language.
6. Risks and guardrails
Wrong language is worse than none: no unreviewed lyrics are ever sung. Languages without a reviewer stay waitlisted.
Cultural respect: original songs only; no copying of traditional lullabies; reviewers can veto imagery that feels off.
Children's data: baby names and recordings are private by default, never used to train models, deletable in one tap.
Voice consent: grandparent voice clones only with recorded consent from that grandparent.
7. MVP build order
Brief form and a Claude prompt that returns lyrics, translation and respelling as JSON, for 3 languages.
Reviewer queue page and 3 paid reviewers (Yoruba, Tagalog, Spanish).
Melody plus TTS spoken guide; sung vocal second.
Pronunciation screen with slow playback and a record button.
Share link for grandparents with a reply-with-voice-note button.