Google shipped Gemini 3.5 Transcribe on August 26, 2026 — a dedicated speech-to-text model, not another audio mode bolted onto a chat model. Google's post frames it as turning "raw audio directly into accurate, polished, formatted text," and the headline demo is the model quietly deleting your "ums."
That framing is the story, and it is also the trap. A model that removes filler words, resolves your self-corrections, and infers your intent is not producing a recording of what was said. It is producing an interpretation. For meeting notes that is the whole point. For a deposition it is a defect. This post is about which of those you are building.
Official Gemini 3.5 Transcribe launch demo, optimized for playback from Google's original X post.
TL;DR — the questions people are actually typing
| Question | Direct answer |
|---|---|
| What are the model IDs? | gemini-3.5-transcribe (pre-recorded, Interactions API) and gemini-3.5-transcribe-live (streaming, Live API). |
| How much does it cost vs Whisper/Deepgram? | |
| Can I stream it? | Yes — gemini-3.5-transcribe-live over the Live API, bidirectional, sub-second latency. Google claims 70% faster time-to-final-transcript than Chirp 3. |
| How many speakers can it separate? | Three, with word-level timestamps. Google labels 3+ speakers experimental. |
| Does it work offline? | No. Hosted API only. Gboard Rambler's on-device behavior is a different, smaller model. |
| Is the filler-word removal optional? | Yes — it is opt-in, not default. Verbatim mode is available. But smart mode is mutually exclusive with diarization and word timestamps. |
| What's the accuracy claim? | 2.6% WER non-streaming, 4.0% streaming (Artificial Analysis); 5.04% / 5.50% on FLEURS. |
| Custom vocabulary? | Up to 1,000 terms; Google recommends staying under ~100 for best results. |
| Where can I use it today? | Google AI Studio and Antigravity (public preview), Gemini Enterprise Agent Platform, the Gemini macOS app, Gboard Rambler on Android. Chrome "coming soon." |
The numbers Google published, and the one the aggregators got wrong
The 2.6% figure that spread across X on launch day is real and it is Google's own. From the launch post: an average word error rate of 4.0% for streaming and 2.6% for non-streaming, measured by Artificial Analysis.
The more useful number is the one fewer people repeated. On FLEURS, the public multilingual benchmark, Google reports 5.04% WER non-streaming and 5.50% streaming. That is roughly double the headline figure, and it is the honest one to quote when someone asks "is it better than what I have," because FLEURS is a shared, public dataset that other vendors also publish against. Artificial Analysis's composite AA-WER is a weighted blend across its own dataset mix — a fine relative ranking, a bad thing to compare against a vendor's self-reported internal number.
The claim that circulated as "distinguishes up to three speakers even in noisy settings" also checks out, with a caveat the reposts dropped: Google's wording is "up to three speakers (support for 3+ speakers is experimental)." Three is the supported ceiling. If your product is a six-person standup recorder, speaker diarization is not solved for you by this release.
| Metric | Gemini 3.5 Transcribe | Source |
|---|---|---|
| WER, non-streaming | 2.6% | Google, via Artificial Analysis |
| WER, streaming | 4.0% | Google, via Artificial Analysis |
| FLEURS WER, non-streaming | 5.04% | Google launch post |
| FLEURS WER, streaming | 5.50% | Google launch post |
| Languages auto-detected | 85+ | Google launch post |
| Time-to-final improvement | 70% vs Chirp 3 | Google launch post |
| Speakers attributed | 3 (3+ experimental) | Google launch post |
| Custom vocabulary terms | Up to 1,000 (≈100 recommended) | Gemini API docs |
Smart transcription vs. verbatim: the section that decides your use case
This is the part of the release that matters most and got the least coverage.
Gemini 3.5 Transcribe has two transcription modes, set through transcription_config. Smart transcription is the one in the marketing: it strips filler words, resolves stuttering and false starts, silently applies your mid-sentence corrections ("let's meet Tuesday — no, Wednesday" becomes Wednesday), and auto-formats the result. Verbatim mode does not. It gives you the words as spoken, disfluencies intact.
Smart mode's output is an interpretation of the utterance, not a record of it. That is a feature when the artifact is a note, a message, a dictated email, or a command. It disqualifies the output entirely when the artifact is evidence:
- Legal — depositions, hearings, recorded interviews. A transcript that resolves a witness's self-correction has destroyed the fact that they corrected themselves, which is frequently the point. Courts have already gotten strict about AI-derived text in filings; explainx.ai covered the sanctions trail from AI-fabricated legal citations, and a silently cleaned transcript sits in the same category of risk.
- Medical — clinical documentation where hesitation, repetition, and self-correction are diagnostic signal, not noise.
- Research — conversation analysis, linguistics, UX research. Disfluency is the data.
- Accessibility captioning — a deaf viewer is entitled to what was actually said, not a tidied paraphrase of it.
Use verbatim for all four. And note the constraint that follows: per Google's transcription docs, smart transcription cannot be combined with word-level timestamps or speaker diarization. You get clean prose, or you get structured multi-speaker output with offsets. Not both in one call.
That single line reshapes a lot of product architecture. A meeting-notes product that wants "who said what, when" and readable prose has to either run verbatim and clean it downstream, or make two passes. Running a small local normalizer after ASR — the Superwhisper S1-mini approach explainx.ai covered in August — suddenly looks less like a niche trick and more like the standard answer, because it decouples "recognize accurately" from "make readable" so you keep timestamps on the first stage.
Function calling inside a transcription model
Logan Kilpatrick's launch note listed function calling among the features, which is unusual enough in an ASR model to be worth unpacking. In Google's framing the model "delegates complex tasks (such as image generation and file analysis) to other Gemini models," and it is live today in the Gemini macOS app.
The conventional voice-agent stack has three network hops:
audio ──▶ ASR ──▶ text ──▶ LLM ──▶ tool call ──▶ execute
(hop 1) (hop 2)
With function calling in the transcription model, the intent-recognition hop collapses into the recognition hop:
audio ──▶ Gemini 3.5 Transcribe ──▶ tool call ──▶ execute
(hop 1)
What that buys you is real but narrower than it sounds:
- Latency. You delete an entire model round-trip from the critical path for utterances that are unambiguously commands. On a voice agent where perceived responsiveness is the whole product, one fewer hop is worth more than a WER decimal point.
- Cost. You stop paying separately for an LLM pass whose only job was to read a transcript and decide it was a tool call.
- Fidelity. The model deciding the tool call has heard the audio, not just read a transcript of it. Prosody and hesitation survive to the decision point instead of being flattened into text first.
What it does not buy you is a reasoning layer. This is intent routing, not planning — do not read it as "you no longer need an orchestrator." Anything with multi-step state still needs the model behind it, and the architecture patterns in the Hugging Face speech-to-speech voice agent guide still apply; you are removing one stage, not the pipeline.
Pricing, honestly
Google's pricing page lists both models in tokens with a per-minute annotation:
| Model | Audio input | Text output | Blended, roughly |
|---|---|---|---|
gemini-3.5-transcribe | $2.00 / 1M tokens (~$0.003/min) | $12.00 / 1M tokens (~$0.002/min) | ~$0.005/min |
gemini-3.5-transcribe-live | $3.50 / 1M tokens (~$0.005/min) | $21.00 / 1M tokens (~$0.004/min) | ~$0.009/min |
Two things to flag. First, the per-minute figures are Google's own conversion at an assumed token rate — your real bill is token-metered, so dense, fast, multi-speaker audio costs more per wall-clock minute than a slow monologue. Budget from tokens, not from minutes.
Second, this is not the cheap option. Published 2026 rates put AssemblyAI's batch tier around $0.0025/min and Deepgram Nova-3 batch around $0.0043/min — both under Gemini's batch price. On streaming, the picture flips: ElevenLabs Scribe v2 Realtime and AssemblyAI Streaming both sit in the $0.39–$0.45/hour range (roughly $0.0065–$0.0075/min), so Gemini's ~$0.009/min live tier is competitive-ish but not a bargain. OpenAI's GPT-Transcribe lineage lands around $0.0045–$0.006/min for batch, which explainx.ai covered at the GPT-Transcribe and GPT-Live-Transcribe launch in July.
If your decision is purely cost-per-minute on English batch audio, Gemini 3.5 Transcribe does not win. You are paying for the smart mode, the function calling, and the 85-language auto-detect.
Should you switch? A decision table
| Your use case | Switch? | Why |
|---|---|---|
| Meeting notes / summaries | Yes, cautiously | Smart mode is genuinely built for this. But you lose diarization when you enable it — so either run verbatim + downstream cleanup, or accept unattributed prose. |
| Voice agents / command interfaces | Yes | Function calling in the recognition hop plus sub-second streaming is the strongest argument in this release. This is the use case Google built it for. |
| Subtitles / captions | No, not yet | Word-level timestamps exist, but smart mode is off the table if you need them, and the 30-minute cap when timestamps are enabled means chunking long-form video. Incumbents are more mature here. |
| Compliance / verbatim / legal / medical | Only in verbatim mode, and audit it | Verbatim mode is the correct setting and it exists. Verify on your own audio that no cleanup leaks through before it touches anything discoverable. |
| Low-resource languages | Test it | 85+ languages with auto-detection is a real advantage over English-first incumbents, and 5.04% on FLEURS is a credible multilingual number. But "supported" and "good enough for your dialect" are different claims — benchmark your own. |
| Privacy-sensitive / on-prem | No | Hosted only. If audio cannot leave your network, this release changes nothing for you — the local Whisper and Parakeet stack is still the answer. |
| India-first / heavy code-mixing | Benchmark against local players | Code-switched Hindi-English is where India-specific stacks like Sarvam's voice agents have earned their position. Do not assume 85 languages means 85 good ones. |
Honest limitations
- Public preview, not GA. AI Studio and the Gemini Enterprise Agent Platform are both preview surfaces. Model IDs, config field names, and per-token prices can all move before general availability.
- The two modes are mutually exclusive with structure. Smart transcription rules out diarization and word timestamps. This is the single biggest architectural constraint in the release.
- Duration caps. Standard requests handle up to about an hour; enable diarization or word-level timestamps and that drops to roughly 30 minutes. Long-form podcast and video workflows need chunking, and chunk boundaries are where speaker labels drift.
- Three speakers is thin. Google's own wording caps reliable attribution at three and calls anything beyond experimental. Most real meetings have more than three people in them.
- No independent verification yet. Every accuracy number in this post is Google's, including the Artificial Analysis attribution. There is no third-party bake-off on 3.5 Transcribe as of publication.
- No offline story. Gboard Rambler on Android does clean up dictation locally, but that is a Gemini Nano-class model, not this API. Do not conflate the two — the Rambler experience is not evidence that 3.5 Transcribe runs on a phone.
- Chrome is "coming soon." Talk-to-type in any web field is the most interesting consumer surface named in the announcement and it has no date.
The takeaway
Gemini 3.5 Transcribe is the first speech-to-text release in a while that is clearly designed around voice as an interface rather than voice as an archive. Function calling in the recognition hop, intent capture, self-correction handling, 85-language auto-detect — every one of those is a voice-agent feature, and they are why you would pay above the batch-transcription market rate.
The thing to get right before you ship is the mode. Smart transcription is not a quality setting, it is an editorial decision about whether your output is a record or a rendering. Pick verbatim any time someone downstream might reasonably expect the former — and check, on your own audio, that verbatim really is verbatim.
Related on explainx.ai:
- Superwhisper S1-mini: on-device transcript cleanup — the local counterpoint to hosted smart transcription
- OpenAI's GPT-Transcribe and GPT-Live-Transcribe — the closest direct rival on price and streaming
- Meetily: privacy-first local transcription — what to use when audio cannot leave the machine
- Sarvam Voice Agents go public — the code-mixed Indian-language benchmark to beat
- Bulbul V4: Sarvam's expressive TTS — the other half of any voice stack
- Hugging Face speech-to-speech voice agents — where a transcription model with tool calls fits in the pipeline
- Cartesia Sonic 3.6 and the voice leaderboard — how Artificial Analysis scores voice models
- Google SL2T brings ASL sign-to-text to Pixel 11 — Google's other 2026 transcription bet
- Gemini 3.7 Flash vs Grok 4.6 vs Sonnet 5 vs GPT-5.6 — the reasoning models behind your voice agent
- Gemini hit 1 billion users — the distribution that makes Gboard and Chrome rollouts matter
Official: Google's launch post · Gemini API audio transcription docs · Gemini API pricing
Model IDs, WER figures, pricing, speaker limits, and duration caps reflect Google's public preview materials as of August 27, 2026 and may change before general availability. All accuracy numbers are Google's own; no independent third-party evaluation of Gemini 3.5 Transcribe was available at publication.