Gemini 3.5 Transcribe: a $0.005/min speech-to-text API, 85+ languages
Real-time and batch from one model, diarization, word timestamps, custom vocab, and a claimed 2.6% WER — live now via Google's Live and Interactions APIs.
Copy markdown
Two model IDs — one for live, one for files
`gemini-3.5-transcribe` handles pre-recorded audio up to an hour per request; `gemini-3.5-transcribe-live` streams with sub-second latency for voice agents. Google says the pair runs about 70% faster than its old Chirp 3 stack.
The price: about half a cent a minute
Pre-recorded transcription lands at ~$0.005/min blended ($2/M audio-input tokens, $12/M text-output tokens); the live streaming variant runs ~$0.009/min. That undercuts most managed speech APIs for high-volume call and meeting workloads.
2.6% WER — but Google graded its own homework
Google reports 2.6% word error rate non-streaming and 4.0% streaming (5.04%/5.50% on multilingual FLEURS). The catch: these are vendor-selected numbers, and nobody outside Google has run an adversarial head-to-head against Whisper or Parakeet yet.
85+ languages, auto-detected, mid-sentence
It detects and transcribes over 85 languages with no language flag and handles code-switching inside a single utterance. It also strips filler words and cleans up self-corrections, so the transcript reads like edited text rather than a raw dump.
Diarization, timestamps, and a 1,000-term vocab
Speaker diarization covers up to 8 speakers (3+ is still experimental), plus word-level timestamps and custom-vocabulary biasing for up to 1,000 jargon terms or unusual spellings. Turning on diarization or timestamps caps input at 30 minutes instead of 60.
Weekend build: a call-to-memo pipeline
Point the Interactions API at recorded sales or support calls, get diarized transcripts with timestamps, then feed them to a summarizer for searchable notes and action items. A post-call analytics tool that used to need Whisper plus a separate diarization library is now two API calls.