User feedback: Scribe consistently produces bad transcripts for Slovenian
narodno-zabavna (NZ) folk-pop music:
- 'Saša Avsenik - Žena ME TEPE': hallucinated 'sam sam sam' x14
- 'FEHTARJI - Gorenjska Ljubljena': total hallucination ('finančni moduli')
- 'Ansambel UNIKAT - PA PA': mistranscribed 'mu' as 'vsem'
- 'Ansambel Saša Avsenika - CVETELE SO MALINE': wrong lyrics entirely
Common pattern: all are Slovenian folk-pop with diatonic accordion. Scribe
training data has very little of this genre, so it consistently fails.
Solution: auto-detect NZ songs by filename keywords and route directly
to Gemini 3 Pro (which handles them correctly), skipping Scribe entirely.
is_likely_folk_pop() detects:
- Slovenian: ansambel, avsenik, slak, fehtar, modrijan, atomik, gadi,
vikend, stil, unikat, korenjaki, gorenjski, štajerski, polka, valček
- Croatian: klapa, thompson, mate bulić
- Serbian/Bosnian: lepa brena, ceca, halid bešlić
When detected:
1. Skip Scribe entirely (it would fail anyway)
2. Go directly to Gemini 3 Pro (~100s, /bin/sh.20)
3. If Gemini fails, fall back to Scribe (rare)
Cost analysis (10 reels/day, 30% NZ):
- Before: 10x Scribe = $0.13/day, ~30% need re-process
- Hybrid (fallback): 10x Scribe + 3x Gemini retry = $0.79/day
- NZ-routing (now): 7x Scribe + 3x Gemini = $0.69/day, FIRST-TRY success
Saves time AND money for NZ-heavy workloads.
Sebastjan Artič
91caf957f2