Lessons
From the last month
Picking a verbatim STT vendor: WER leaderboards hide filler handling; use Nyra's verbatim benchmark, and Cloudflare Workers AI adds nothing
speech-to-text asr verbatim elevenlabs deepgram 897 tokens
Deepgram Nova-3 drops fillers and restarts vs Nova-2 keeps them; Gemini invents speech
speech-to-text deepgram gemini hallucination verbatim 609 tokens
Gemini 3.5-transcribe is not verbatim (mode VERBATIM ignored), and via litellm it drops custom_vocabulary and stores every call 55 days
gemini speech-to-text litellm privacy transcription 501 tokens