cookbook/data_labeling/_11_audio_transcription/TEST_LOG.md
Tested 2026-07-18 against gemini-3.5-flash, agno 2.7.4.
Status: PASS
Description: Verbatim transcript of QA-01.mp3 (English family Q&A clip) into a flat Transcript schema with a single text field.
Result: Full clean transcript returned, starting "How many people are there in your family? There are five people in my family..." and ending "...My mom always prepares delicious meals for us." Tokens: input=1949, output=1488; duration 8.23s.
Status: PASS
Description: Transcript of sample_conversation.wav split into speaker turns via the DiarizedTranscript schema, with generic speaker identifiers.
Result: Five turns returned alternating between exactly two speakers, labeled consistently as "Speaker A" and "Speaker B" (A opens with "Hello, Liam, hey..." and closes with "...Thanks for the encouragement."). No invented names. Duration 3.38s.
Status: PASS
Description: Transcript of QA-01.mp3 split into sentence-level segments with start/end times in seconds via the TimedTranscript schema.
Result: 25 segments returned from 0.0 to 116.0 with monotonically non-decreasing times, e.g. [0.0, 2.1] "How many people are there in your family?". Known model quirk observed: past the one-minute mark the model flattens mm:ss into decimal (59.3 jumps to 103.3, i.e. 1:03.3 read as 103.3 seconds), so absolute values after 60s are inflated while ordering stays correct.