Build with Beam · hackathon entry

Singalong

Sing a foreign song in your own (silly) words. Pick one of 40 public-domain songs from every region of the world, or paste lyrics (never stored), choose the language you sing in, and Beam rewrites each line into words you know that sound like the original. The real meaning stays underneath.

Tap a line to hear the original and then your version. Matched syllables light up; mismatches dim. Then sing it karaoke-style and make a clip.

Demo: pick a song, watch Beam rewrite it, sing karaoke

Some real outputs

Kalinka (Russian) → English
Call a king, call a king, call a king, my oh!
O Tannenbaum (German) → English
Oh, tanning balm, oh tanning balm
Ochi chyornye (Russian) → Spanish
Ocho churros, ocho estrés
Jana Gana Mana (Hindi) → Indonesian
Jangan gatal mana, adik naik jaya, hei!

How Beam is used

  1. Code: lyrics → IPA (phonemizer + espeak-ng; epitran for Devanagari).
  2. Beam drafts the verse from the lines and their IPA as one continuous silly story, clean only.
  3. Code measures each line: phonetic closeness, syllable count, stress, share of real target-language words, and whether it just copies the source.
  4. Beam fixes only the failing lines, given the measured reason, keeping the story (≤2 rounds; a rewrite is kept only if it measures better).

Evals (honest numbers)

All 40 songs × 2 target languages = 80 runs, 320 lines.

DraftAfter fix loop
Lines passing all code checks23.7 %38.1 %
Phonetic score (mean)0.6660.696
Real target-language words97.9 %99.0 %
Lines copying the source1.3 %0.3 %

By target language (pass rate draft → final): English 31 → 51 %, Indonesian 25 → 40 %, French 21 → 33 %, Hindi 21 → 33 %, Portuguese 21 → 29 %, Spanish 17 → 25 %, Italian 17 → 21 %, German 4 → 11 %. On an earlier easy 4-song set the loop reached 88 %; the world corpus (Korean, Xhosa, Hawaiian, Arabic…) is much harder.

Error analysis (agent-led review of traces, not a human expert): the main failure is syllable salad, real words that only echo each syllable ("Sew nay lay ma teen"). Others: translation or cognate copies ("Prospero is the wind"), Hindi that only transliterates, and a field-echo bug (fixed). Copies and non-words are now code checks.

LLM judge for syllable salad: validated against 80 agent-labelled lines. Dev TPR 0.65 / TNR 0.90; held-out test TPR 0.48 / TNR 0.84. That is not good enough to trust, so the app doesn't use it. Human labels are the next step.

Rules I kept

Public-domain demo songs only, with source and licence for each (corpus table). Pasted lyrics are never stored or logged. Clean-only prompt + code filter. Beam spend capped per day on the server. Total Beam spend to build and evaluate: about $4.8.