Google opens Gemini 3.5 Transcribe for real-time and recorded speech APIs

I want to know how well it handles proper subtitle formatting. I am learning languages, and film or video transcripts would be useful for checking my understanding.

I do not see it in Gboard yet even though the post says it is available on Android. I assume this is another gradual Google rollout.

I think the naming is confusing. Rambler appears to be a Gboard subfeature limited to Pixel 10 and newer Samsung phones, and users just see improved transcription rather than a separate app or switch.

WER misses my biggest problem with transcription: random sentence breaks that make the text obviously wrong. Fixing those on Android can take longer than typing, and I often notice errors only later.

I would run another model over the transcript for cleanup. A local STT demo did that with a clean-transcript.md full of model-specific gotchas, and it fixed many errors.

I want to know whether it hallucinates like Chirp. Chirp can turn noise or silence into repeated text for minutes, so we went back to Whisper for timestamps and Gemini Flash for transcription correction.

I use it all day for dictation, and it is the best model I have used for accuracy, latency, and formatting.

Whisper can also repeat words near the end and overwrite actual content. It is otherwise very good, but noisy car audio exposes that failure mode.

I tested it heavily with silence and noise because hallucinated output is unacceptable for us. I got no hallucinations, though exactly 20 seconds of silence can randomly produce a 403.

I tested it on a Pixel 11 Pro and dislike how Smart transcription simplifies wording. It repeatedly changed "I hesitated to check it, I should have verified" into "I should have verified," removing meaning I intended to keep.

Are you using Smart or Verbatim mode? The API documentation distinguishes the two transcription modes.

I benchmarked more than 20 STT models on German, Italian, and English meeting audio with industry terms and language switching. Voxtral Mini 3B is the only local model that satisfies me, and ElevenLabs is the only paid API that is slightly better; I would still prefer something that runs on my machine.

I agree that Voxtral handles multilingual conversation unusually well. Mixed-language sentences and industry terms do not seem to bother it as much as other models.

I assume all the compute happens in the cloud.

I have been using Wispr Flow on my Pixel 9 and like it. It already does what Rambler claims to do, while Rambler is limited to Pixel 11 for now.