Gemini 3.5 Transcribe lands on the API: 85+ languages, diarization
Speech-to-text is having a week: Google's ASR endpoint arrives as Fish Audio undercuts on price and Grok's transcriber holds the accuracy crown.
Copy markdown
Google ships a dedicated ASR endpoint, 85+ languages
Gemini 3.5 Transcribe is a speech-to-text model on the Gemini API — word-level timestamps, speaker diarization up to 8 speakers, custom vocabulary (1,000 terms), and filler-word cleanup, with auto language detection across 85+ languages. It's about $2/M input and $12/M output, exposed as both a unary (gemini-3.5-transcribe) and streaming (-live) endpoint; audio up to an hour, or 30 minutes with timestamps or diarization enabled. You can drop a real transcription layer into an app today without self-hosting Whisper.
Fish Audio undercuts on price for long-form audio
Transcribe 1 Pro bills at $0.0001 per second — about $0.36 per hour of audio — and is tuned for interviews, meetings, and podcasts: speaker labels, word-level timestamps, auto language detection, and it preserves emotion and vocal-event cues. For anyone transcribing hours of recordings, per-second billing is far easier to reason about than per-token.
The bar to beat: Grok Voice Transcribe 2.0 at 2.7% WER
xAI's transcriber sits at the top of Artificial Analysis's streaming-accuracy ranking at a reported 2.7% word error rate — roughly double the accuracy of its predecessor at the same price. That's the number this week's new arrivals have to clear, so benchmark on your own audio before you wire any of them into a pipeline.
Elsewhere: Fireworks' Ember-1 trims reasoning tokens 40%
Not speech, but it shipped the same day: Ember-1, a reasoning model built on Kimi K3 that Fireworks says matches base quality while using about 40% fewer reasoning tokens. It's $3/M in, $15/M out, with a 1M-token context, live now on OpenRouter — cheaper agentic runs if that token saving holds on your workload.