Tag
3 posts
An eval across 12 audio-capable models on OpenRouter, 5 MP3 bitrates, 4 dictation samples. The default upload quality in most dictation apps is 2–4× overprovisioned — and the more interesting finding had nothing to do with bitrate.
An MCP server that turns audio files into transcripts, meeting notes, blog posts, emails, and dev specs using Google Gemini — with 200+ transformation presets, VAD preprocessing, and SSH file retrieval.
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.
