Open Source Benchmark

Gemini Transcription: Amazing Accuracy,
Timestamps Gone Wrong

WER improves ~1%: no-think 5-6.5% → think 4.3-4.8%
DER unchanged: Both modes hover around 30-33% for Flash
Pro is unstable: think-URL fails (311%), no-think-Local fails (306%)
Critical insight: We examined the thinking output—it's just a draft transcript , not timestamp reasoning

We systematically tested Gemini's transcription across 4 dimensions. The text is excellent—but the timestamps need help.

30-300%
Original DER

→

9-16%
+LattifAI DER

Test Material: OpenAI GPT-4o Launch (~26 min) — 4 speakers including ChatGPT, frequent interruptions, overlapping speech, audience applause. Each experiment runs 2+ times for stability verification.

Gemini Transcription: Amazing Accuracy,Timestamps Gone Wrong

Gemini Transcription: Amazing Accuracy,
Timestamps Gone Wrong