Gemini-3.5-Transcribe

Who Should Read

A backend developer choosing an STT model for a voice agent or call-analytics pipeline.

Core Mechanics

• Separate APIs for streaming and recorded audio
  • Real-time uses the Live API with gemini-3.5-transcribe-live.
  • Recorded audio uses the Interactions API with gemini-3.5-transcribe.
  • The recorded-audio API supports speaker attribution and word-level timestamps.
• It cleans up self-corrections and filler words
  • It handles corrections like "Tuesday—no, Wednesday".
  • It removes ums and ahs and auto-formats the text.
  • It adapts transcription to a custom vocabulary you provide.
• WER is 4.0% streaming and 2.6% non-streaming
  • These are average WER figures measured by Artificial Analysis.
  • On FLEURS across top languages, WER is 5.50% streaming and 5.04% non-streaming.
  • Google says it improves on Chirp 3 on FLEURS.
• Latency is much lower than Chirp 3
  • Time to final transcription improves 70% per Artificial Analysis.

Community Reactions

• Rankings change on your own data, per one tester
  • One commenter tested 20+ STT models on German, Italian and English company meetings.
  • Only Voxtral Mini 3b locally and ElevenLabs among paid APIs satisfied them.
• Accurate but slow, some commenters say
  • A real-time translator builder rated Soniox STT v5 best.
  • They said Gemini 3.5 Transcribe beats others on accuracy but needs better latency.
  • Another user's benchmark site shows it isn't at the frontier on latency or accuracy.
• Filler cleanup can delete meaning, per one user
  • On a Pixel 11 Pro, the first clause of "I hesitated to check it, I should have verified" was dropped every time.
  • They said this is a problem when precise wording matters.
• Chirp hallucinated on noise, one user reports
  • Noisy or silent audio made it repeat a phrase until request timeout.
  • They moved back to Whisper for timestamps and Gemini Flash for correction.

How to Apply

• Benchmark on 10 of your own samples
  • Run 10 real call or meeting recordings through gemini-3.5-transcribe in AI Studio today.
  • Compare side by side with your current STT and count errors in names, numbers and language switches.
• Split real-time and batch by use case
  • Use the Live API with gemini-3.5-transcribe-live for latency-critical voice agents.
  • Use the Interactions API with gemini-3.5-transcribe for call analytics needing accuracy and speaker labels.
• Add domain terms as custom vocabulary
  • List product names, acronyms and order-ID formats and pass them with the request.
  • Compare error rates before and after on the same samples.
• Check whether cleanup removes meaning
  • Do not trust cleaned transcripts as-is in legal or audit pipelines.
  • Sample raw versus cleaned output to catch dropped clauses.

Terminology

Related Papers

• The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge

  Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.

• 90% of US working hours are spent on tasks not covered well by AI benchmarks

  Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage

• Postgres 19 delayed after AI reviewers find more bugs than humans can fix

  LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted

• Show HN: Gemini 4 Argon plays StarCraft

  LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.

• A Plausible Explanation Is Enough to Mislead Multimodal Judges

  Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.

• Neural networks are clean

  Looking at neurons directly shows most have one main job, challenging superposition.

Related Resources

• https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/
• https://ai.google.dev/gemini-api/docs/pricing#gemini-3.5-tra...
• https://multilingualsttbench.com/
• https://fliptalk.ai