Tag

evaluation

3 posts

Mar 26, 2026

Testing Gemini 3.1 Flash Lite's Audio Understanding With 49 Structured Prompts

I built a 49-prompt test suite to evaluate Gemini 3.1 Flash Lite's audio understanding capabilities across 13 categories — from accent detection to deception analysis. Here's what worked, what didn't, and why it matters.

AI Gemini audio multimodal

Mar 25, 2026

Benchmarking speech-to-text on long-form audio

Comparing 8 STT models on a 27-minute podcast. Local Whisper wins on word accuracy, but cloud APIs dominate punctuation.

Projects AI Open Source Whisper

Mar 25, 2026

Does fine-tuning Whisper actually improve accuracy?

Evaluating whether fine-tuning Whisper improves transcription accuracy. Spoiler: it depends on model size and use case.

Projects AI Open Source Whisper