# Expressive Ukrainian audition: machine listening accepted

Three complete phrases, 16 word occurrences, 12.528-second MP3. A separate native audio-analysis pass received the actual audio without the intended transcript, then returned every heard word and lexical stress. The returned sequence matches the intended ledger, including every repeated «Мамо»; all 15 multisyllabic words match their intended stress.

- Opening: questioning rise and concerned curiosity.
- Reassurance: softer, lower register and intimate conviction.
- Resolution: brighter, welcoming delivery with a smile in the voice.

No music or spoken direction tags were reported. This is a short provider experiment, not a replacement for the full film or human native-speaker approval. Returned timings are model estimates, not independently forced-aligned. Owner feedback can still reject the delivery.

[Exact request](request.json) · [Occurrence ledger](pronunciation.json) · [Machine listening record](listening.json) · [Raw blind observations](../../analysis/evidence/13-expressive-audition-v5-blind/analysis.md).

Earlier attempts are retained: the original unsupported voice failed input validation; the supported ElevenLabs v3 attempt failed at the provider. Gemini with three separate turns returned only the opening seven words and was rejected. Combining all three directed phrases in one turn recovered the complete recording. Provider WAV bytes are retained separately; the final audition is normalized to MP3 for browser playback. No failed attempt was silently approved.