FULL STORY

Gemini 3.5 Transcribe Launch Sparks Open-Source Follow-ups

Google launched Gemini 3.5 Transcribe on Aug 27, and developers quickly built open-source dictation apps on top of the new speech-to-text model.

2026-08-27 ~ 2026-08-28 · 2 episodes · 31 posts

Episode 1 · Google Launches Gemini 3.5 Transcribe Speech-to-Text Model (2026-08-27, 29 posts)

On Aug 27, Google launched Gemini 3.5 Transcribe, announced by Sundar Pichai on X. The speech-to-text model emphasizes high-accuracy transcription with smart post-processing, shipping in both pre-recorded and real-time streaming (Live) versions, with APIs live on Google AI Studio and Gemini Enterprise and integrated across Google products. Third-party benchmarks are strong though not top-ranked, and community reception is largely positive, with some puzzled by function calling in a transcription model.

Confirmed

  • Supports 85+ languages and regional accent auto-detection, real-time streaming and batch processing, lower WER, sub-second latency, and mixed-language input (e.g., Chinese-English)
  • Speaker identification, speaker diarization, word-level timestamps, and custom vocabulary for jargon and specialized terms
  • Smart post-processing: filters filler words, intent-based rewriting and formatting; per @stefanjblos's relayed test, offers both verbatim and SMART modes
  • Enhanced understanding of phone numbers, postal codes, and order IDs, robust in noisy environments
  • Function calling: with permission, combines screen context and chat history (e.g., filenames, agent reasoning) to execute voice commands, designed for voice interaction in Google Antigravity
  • Two variants: standard for recorded audio (with diarization and word timestamps), Live for real-time voice applications
  • Per Artificial Analysis: non-streaming AA-WER of 2.6% (rank 5), 84x real-time speed, 0.4s latency; Live reaches 5.8% accuracy 0.25s after end-of-speech detection, faster first response than GPT Live
  • @philschmid's tests: non-streaming WER 2.6%, streaming WER 4.0%, 70% faster final transcription vs Chirp 3, with excellent post-processing
  • @GoogleAI's weekly roundup highlighted the model alongside Gemini Omni 1.1 Flash

Unconfirmed

  • @giffmana questioned why a transcription model needs Function Calling; the official design rationale remains unexplained
  • Rank 5 on AA-WER implies a gap to leading models; full leaderboard details were not provided

Why it matters

  • Background noise, filler words, and misspellings are long-standing dictation pain points the model directly targets
  • Speaker diarization, custom vocabulary, and function calling make it embeddable in real-time voice apps and enterprise workflows
  • Developers including @thorwebdom and @philschmid have posted demos and positive hands-on results, signaling real-world potential in professional scenarios

9 more related posts →

Episode 2 · Developer Open-Sources Dictation App Built on Gemini 3.5 Transcribe (2026-08-27, 2 posts)

A developer open-sourced Jot, a dictation app built with Google's Gemini 3.5 Transcribe model, which supports 85+ languages with smart corrections and custom vocabulary.