HN Debrief

Gemini-3.5-Transcribe

  • AI
  • Developer Tools
  • Mobile
  • Open Source

Google posted Gemini 3.5 Transcribe as a new speech-to-text model aimed at both API use and Android dictation. The pitch is not just lower word error rate, but cleaner formatting and a more polished dictation experience. People reading it immediately separated those two jobs. For straight transcription, the big question was whether it beats the current mix of Whisper, ElevenLabs, and newer local models on noisy audio, timestamps, multilingual speech, and jargon. For dictation, the concern was different. People want exact wording, not a model that tidies up what they said and quietly changes meaning.

If you buy or build around speech input, test for your actual failure modes instead of headline benchmark scores. Mixed-language audio, domain terms, silence handling, punctuation, and whether the model paraphrases your words will decide whether this is usable in production.

Discussion mood

Cautiously positive with a lot of caveats. People liked the apparent jump in accuracy, latency, and noise handling, but many doubted it was enough to displace existing setups because exact wording, multilingual jargon, price, timestamps, rollout confusion, and cloud-only deployment still look like weak spots.

Key insights

  1. 01

    Mixed-language meetings remain the hard test

    Real business audio still breaks many speech systems in ways benchmark numbers hide. The most credible example was meetings that switch between German, Italian, and English while dropping in specialized terms. Voxtral Mini 3B was singled out as the only local model that consistently held up there, and Parakeet 3 was said to fail by bending jargon into words from the dominant language. That is a much tougher bar than clean single-language clips.

    If your users code-switch or use niche terminology, build your eval set around that before considering a vendor switch. A model that tops public leaderboards can still be worse on the audio that matters to your company.

      Attribution:
    • Lucasoato #1 #2
    • hinnisdael #1
  2. 02

    Dictation fails when the model rewrites intent

    For dictation, users are not grading a transcript. They are delegating exact wording. One tester said the Pixel experience repeatedly dropped a clause like "I hesitated to check it" and kept only the cleaner conclusion. That makes the output more readable but less faithful. It turns transcription into summarization, which is the wrong product behavior when someone is composing a message.

    Test dictation systems for semantic fidelity, not just readability. If your app involves legal, sales, support, or sensitive messaging, add checks for dropped hedges, qualifiers, and corrections.

      Attribution:
    • Crystalin #1
    • sync #1
  3. 03

    Silence and tail-end failures still decide reliability

    The ugly production failures are not obvious recognition errors. They are what the model does when nothing useful is there, or when audio is ending. Chirp was criticized for hallucinating text on noise or silence. Whisper was criticized for repeating words near the end and overwriting real content. Gemini 3.5 Transcribe got credit for not hallucinating in silence tests, but one evaluator still found a reproducible 20-second silence bug that returned a 403. That is a better failure mode than fabricated text, but it still needs guardrails.

    Add explicit tests for silence, background noise, truncated endings, and long pauses. Production speech systems need rejection behavior and error handling as much as they need high recognition scores.

      Attribution:
    • film42 #1
    • sync #1
    • wolvoleo #1
  4. 04

    Post-processing is becoming part of the core stack

    Recognition quality is close enough that cleanup now matters a lot. People were more frustrated by bad sentence breaks and awkward punctuation than by individual misheard words. The proposed fix was not "wait for a better STT model" but run a second pass that repairs transcript structure, either with the same model or with a specialized cleanup model like superwhisper/s1-mini. That is a strong sign the product boundary is shifting from raw transcription to transcript editing pipelines.

    Plan for a two-stage architecture if transcript quality affects user trust. Budget for punctuation, formatting, and consistency cleanup instead of assuming the speech model alone will solve the user experience.

      Attribution:
    • jeffbee #1 #2
    • verdverm #1
    • coder543 #1

Against the grain

  1. 01

    Price and quality still favor ElevenLabs

    Not everyone thought Google had reached the front of the pack. One commenter flatly said Gemini 3.5 Transcribe is both more expensive and worse than ElevenLabs Scribe, which cuts against the idea that Google has a broadly superior offering. Even if that view is narrow, it matches the wider skepticism that this launch does not yet beat incumbent tools on the dimensions practitioners care about.

    Do side-by-side cost and quality comparisons before migrating from an existing vendor. Google's distribution advantage does not automatically make this the best API choice.

      Attribution:
    • dbbk #1
  2. 02

    Cloud-only deployment is a real blocker

    The model runs in the cloud, and that alone rules it out for some buyers. People already happy with local options like Voxtral Mini 3B were not arguing about tiny quality deltas. They were pointing out that easy on-device deployment, privacy control, and avoiding API dependence can outweigh a modest accuracy gain.

    If your product handles sensitive audio or needs predictable cost and latency, keep local models in the shortlist. A cloud-only win has to be large enough to justify the operational tradeoff.

      Attribution:
    • hypfer #1
    • k9294 #1
    • Lucasoato #1

In plain english

403 ↩
An HTTP status code meaning access is forbidden, used here to describe an API error response.
API ↩
Application Programming Interface, a defined way for software to expose functions or data to other software.
Chirp ↩
Google's earlier speech recognition model family, mentioned here as a comparison point.
ElevenLabs Scribe ↩
A commercial transcription product from ElevenLabs mentioned as a benchmark and competitor.
Gboard ↩
Google's keyboard app for Android, which includes voice input and dictation features.
Parakeet 3 ↩
A speech recognition model mentioned by commenters as fast but weaker on mixed-language business jargon.
Pixel ↩
Google's line of Android phones.
STT ↩
Speech-to-text, the process of converting spoken audio into written text.
superwhisper/s1-mini ↩
A Hugging Face-hosted model mentioned as a possible second-pass cleanup tool for fixing punctuation and formatting in transcripts.
Voxtral Mini 3B ↩
A small speech model mentioned in the comments that can run locally, with 3B referring to roughly three billion parameters.
Whisper ↩
OpenAI's speech recognition model family, widely used for transcription and subtitles.
Word Error Rate ↩
A common speech recognition metric that measures how many words in a transcript were inserted, deleted, or substituted compared with a reference transcript.

Reference links

Google product docs and rollout

Alternative models and tools