Gemini 3.5 Transcribe: Google launches an 85-language STT engine at 2.6% WER, 70% faster than Chirp 3 — and it auto-cleans hesitations
🔎 Why a dedicated STT model changes the game
Until now, speech-to-text was a problem we solved with general-purpose models cobbled together for transcription. Whisper, Deepgram, Chirp 3: they all get the job done, but none were designed exclusively to turn voice into clean text.
Google just changed the game with Gemini 3.5 Transcribe, announced in public preview on August 26, 2026. This isn't a chatbot that knows how to transcribe. It's a specialized STT model, with two distinct endpoints (batch and streaming), that achieves a 2.6% Word Error Rate in batch — a score that rivals human performance on certain tasks.
The detail that hits the mark: it automatically removes filler words, hesitations, and repetitions. Your transcript comes out clean, ready to be published or analyzed. No more post-processing needed.
Another strong signal: Google is gradually phasing out generateContent as the default interface for Gemini agents in favor of the Interactions API. Gemini 3.5 Transcribe is one of the first models to fully leverage this new architecture.
Key takeaways
- Two specialized endpoints: batch (Interactions API, 2.6% WER) and live streaming (4.0% WER), each optimized for its use case.
- 85+ languages with automatic detection and code-switching handling (switching languages mid-sentence).
- Smart transcription: the model filters out filler words ("um", "like", "well") and repetitions — the final text is ready to use.
- Aggressive pricing: ~$0.005/minute in batch, free tier on AI Studio.
- Immediate integrations: LiveKit, Pipecat, Agora, Fishjam, Vercel AI Gateway. Gboard deployment, Gemini macOS app, and Chrome coming soon.
- Successor to Chirp 3, 70% faster on time-to-final-transcript. API only, no open weights.
Recommended tools
| Tool | Main use | Price (August 2026, check on ai.google.dev) | Ideal for |
|---|---|---|---|
| Gemini 3.5 Transcribe (batch) | Asynchronous transcription | $2 / $12 per 1M tokens (~$0.005/min) | Podcasts, meetings, interviews |
| Gemini 3.5 Transcribe (live) | Real-time transcription | $3.50 / $21 per 1M tokens | Live calls, subtitles, voicebots |
| AI Studio | Free testing | Free tier included | Rapid prototyping, POCs |
| LiveKit | Live audio infrastructure | Variable | Real-time voice integration |
| Vercel AI Gateway | AI routing and cache | Variable | Next.js apps with transcription |
Two endpoints, two philosophies
Gemini 3.5 Transcribe does not offer a single model with a "fast" mode and a "precise" mode. Google designed two truly distinct endpoints, each with its own optimization.
Batch: precision above all
The batch endpoint uses Google's Interactions API — the new default interface for Gemini agents. You send an audio file (up to 1 hour per request), you wait, and you receive a transcript at 2.6% WER.
This 2.6% WER score in batch is measured by Google on internal benchmarks. To give you an idea, Whisper Large v3 hovers around 5-7% WER on conversational English according to independent benchmarks. We are in a completely different league.
The obvious use case: recorded meetings, podcasts, interviews. Anything that is not real-time and where transcript quality is paramount.
Live streaming: uncompromising responsiveness
The live endpoint achieves 4.0% WER in streaming. Logically, this is not as good as in batch — the model does not have the complete context of the sentence. But 4.0% in real time is excellent.
The key advantage: the time to obtain the final transcript is 70% faster than with Chirp 3, Google's previous model. For live subtitles or voice assistants, this reduced latency makes all the difference.
Smart transcription: the transcript that cleans itself
This is perhaps the most underrated feature of the announcement. Gemini 3.5 Transcribe includes a "smart transcription" that automatically filters out irrelevant elements.
Specifically, the model identifies and removes filler words ("uh", "um", "like"), hesitations ("I'm going to say... I mean..."), and involuntary repetitions. The result contains only the structured speech.
The Decoder, which analyzed the model in detail, emphasizes that this ability to self-correct verbal stumbles is integrated natively — not added via a prompt or separate post-processing. It is the model itself that decides what to keep and what to remove.
For content producers, journalists, and legal teams: this eliminates an entire step in the workflow. No more need to proofread the raw transcript to clean it up before publication or analysis.
85 languages, code-switching and diarization
Automatic detection and code-switching
The model supports 85+ languages with automatic detection. You don't need to specify the input language — the model detects it on its own.
Even more interesting: native code-switching. That is, the ability to correctly transcribe a speaker who changes languages mid-sentence. A French-Moroccan alternating between French and Moroccan Arabic, an Indian switching from Hindi to English — the model handles it seamlessly.
For international companies and multilingual markets (Africa, India, Southeast Asia), this is a major asset. None of the direct competitors handle this scenario as well natively.
Diarization: who said what?
Diarization is available for up to 3 speakers in the stable version, and up to 8 speakers in experimental mode. Every word in the transcript is associated with a speaker identifier.
Coupled with word-level timestamps, this makes it possible to accurately reconstruct the timeline of a conversation: who spoke, when, and exactly what they said.
For smart meeting tools, this is the winning combo. No need for a separate diarization model — everything is integrated into a single API call.
Detailed pricing: Google attacks on price
The pricing of Gemini 3.5 Transcribe is structurally aggressive. Here is the comparison table (August 2026, check on ai.google.dev):
| Endpoint | Input (1M tokens) | Output (1M tokens) | Estimation per minute |
|---|---|---|---|
| Batch | 2 $ | 12 $ | ~0.005 $ |
| Live streaming | 3.50 $ | 21 $ | ~0.009 $ |
MLQ.ai calculated a blended cost estimate of around 0.005 $/minute in batch. This is extremely competitive. For context: Deepgram generally charges between 0.004 $ and 0.012 $/minute depending on the model and volume. Self-hosted Whisper costs less in tokens but more in infrastructure.
The free tier on AI Studio allows you to test without a credit card. This is a clear signal of Google's strategy: to make adoption frictionless for developers.
Comparison: Gemini 3.5 Transcribe vs Whisper vs Deepgram
OrcaRouter published a solid comparative analysis upon the model's release. Here is a summary:
| Criterion | Gemini 3.5 Transcribe | Whisper Large v3 | Deepgram Nova-2 |
|---|---|---|---|
| Batch WER (English) | 2.6% | ~5-7% | ~3-5% |
| Streaming WER | 4.0% | N/A (non-native) | ~5-7% |
| Languages | 85+ | 99 | 30+ |
| Code-switching | Native | Partial | Limited |
| Smart transcription | Yes (native) | No | No |
| Diarization | 3 (8 exp.) | Via external pipeline | 3-5 |
| Word-level timestamps | Yes | Yes | Yes |
| Pricing (~/min) | $0.005 (batch) | Free (self-hosted) | $0.004-0.012 |
| Live latency | 70% faster than Chirp 3 | Not applicable | Good |
| Open weights | No | Yes | No |
Whisper remains relevant if you have data sovereignty constraints (on-premise) or need fine-tuning. But in pure SaaS, Gemini 3.5 Transcribe takes the lead on almost all technical criteria.
To choose between the LLM models that could complement this transcription workflow, our Gemini vs ChatGPT vs Claude comparison remains the reference.
Integrations: ready to use, not a lab toy
A powerful STT model without an integration ecosystem is dead paper. Google has clearly spent time on the partnership side before the launch.
Voice AI frameworks
Gemini 3.5 Transcribe is already integrated with:
- LiveKit: open source audio/video infrastructure, very popular in the voice AI community
- Pipecat: Python framework for building multimodal voice pipelines
- Agora: real-time SDK for audio calls
- Fishjam: SFU framework for real-time applications
- Vercel AI Gateway: intelligent routing and cache for AI calls in Next.js apps
These integrations mean that a developer can plug Gemini 3.5 Transcribe into a voicebot in a few lines of code, without managing the underlying audio infrastructure.
Google products
On the consumer side, the model is already landing in:
- Gboard / Rambler on Android: improved voice transcription
- Gemini app on macOS: dictation and transcription
- Chrome: integration planned in the short term for voice dictation in the browser
This product distribution is a massive competitive advantage. Deepgram and Whisper do not have an equivalent consumer distribution channel. Google can test and improve its model at the scale of billions of users.
Concrete use cases by sector
Media and journalism
A journalist records a 45-minute interview in French with some passages in Arabic. They send the file via the batch API. Within minutes, they receive a clean transcript, free of hesitations, with diarization and timestamps. They can directly copy-paste the quotes into their article.
Smart transcription easily saves 30 to 45 minutes of cleaning per hour of audio. For a newsroom processing 10 interviews a day, that's half a workload recovered.
Corporate meetings
Integrated via LiveKit or Pipecat into a meeting tool, Gemini 3.5 Transcribe can provide real-time summaries with speaker identification. Code-switching is valuable in multinational companies where teams mix local languages and English.
Voicebots and voice assistants
The streaming endpoint at 4.0% WER with a 70% latency reduction compared to Chirp 3 is built for voicebots. An assistant that poorly transcribes a voice query cannot respond to it correctly — STT is the critical link in the chain.
For developers building voice agents with LLMs like Gemini 3.1 Pro or GPT-5.5 on the backend, Gemini 3.5 Transcribe on the input and a TTS on the output form a complete voice pipeline.
Subtitling and accessibility
Word-level timestamps enable word-by-word synchronized subtitling. For live subtitling (streaming, conferences, broadcasts), the reduced latency is a game-changer compared to previous solutions.
Limitations and points of caution
No open weights
Gemini 3.5 Transcribe is API only. You cannot download it, fine-tune it on your data, or deploy it on-premise. For organizations with data sovereignty constraints (banking, defense, healthcare), this is a blocker.
In this case, Whisper remains the only credible open-source option. Or the free AI APIs for less sensitive use cases.
1 hour maximum per request
The 1-hour limit per batch request is sufficient for most cases, but long recordings (3-hour conferences, legal depositions) require cutting into segments. This is a point to manage in your pipeline.
Experimental 8-speaker diarization
The 8-speaker diarization is marked "experimental". For roundtables with many participants, the stable 3-speaker version may be insufficient. It will be necessary to test in real-world conditions before relying on it in production.
WER measured by Google
The WER figures (2.6% batch, 4.0% streaming) are published by Google on internal benchmarks. They are likely reliable, but an independent benchmark (as the HuggingFace team regularly does for Whisper) would be welcome to confirm.
How to get started: quick guide
Step 1: AI Studio
Go to AI Studio (aistudio.google.com), create a free account. The free tier gives access to Gemini 3.5 Transcribe without a credit card. Upload a test audio file, select the model, launch a batch transcription.
This is the fastest way to judge the quality of the smart transcription on your own data for yourself.
Step 2: Batch API call
For the batch endpoint, you use the Interactions API. The flow is simple: authentication via API key, sending the audio file (up to 1 hour), waiting for the result, retrieving the transcript with diarization and timestamps.
The official documentation at ai.google.dev/gemini-api/docs/models/gemini-3.5-transcribe covers the parameters: language (optional, auto-detection by default), enabling diarization, output format.
Step 3: Streaming integration
For real-time, the easiest path is through a partner framework. LiveKit and Pipecat have both published code examples to plug in Gemini 3.5 Transcribe in a few minutes. If you are already using one of these frameworks, it's a model change, not a rewrite.
❌ Common mistakes
Mistake 1 : Confusing specialized STT and generalist LLMs
Gemini 3.5 Transcribe does not answer questions. It transcribes. If you want an agent that listens, transcribes, understands, and answers, you must chain Gemini 3.5 Transcribe (input) with an LLM like Gemini 3.1 Pro or GPT-5.5 (reasoning) and a TTS (output). These are not the same models.
Mistake 2 : Using streaming for batch tasks
The streaming endpoint costs 75% more than batch ($3.50 vs $2 per 1M input tokens) and has a higher WER (4.0% vs 2.6%). If you don't have a real-time constraint, always use batch.
Mistake 3 : Ignoring code-switching in your tests
Many teams test the model on pure English and conclude that "it's just a good STT". The real differentiator is measured on multilingual conversations with language switching. Test with your real data, not Wikipedia samples.
Mistake 4 : Relying on 8-speaker diarization in production
This is experimental. If your use case depends on diarization beyond 3 speakers, wait for the stable version or plan for a fallback.
❓ Frequently Asked Questions
Does Gemini 3.5 Transcribe directly replace Chirp 3?
Yes, it is its official successor. Google reports a 70% reduction in time-to-final-transcript and a significant decrease in WER. Chirp 3 will likely be gradually deprecated.
Can Gemini 3.5 Transcribe be used for free?
Yes, via the AI Studio free tier. The exact limits of the free tier have not yet been detailed, but Google allows prototyping at no cost. For production, token-based pricing applies.
Does the model handle French correctly?
Yes, French is among the 85+ supported languages. Franco-Arabic, Franco-English, or Franco-German code-switching is handled natively, which is a major advantage for international French-speaking use cases.
What is the difference compared to an LLM that transcribes?
A general-purpose LLM (Gemini 3.1 Pro, GPT-5.5) can transcribe audio via a multimodal mode, but this is a side effect. Gemini 3.5 Transcribe is architected for STT: better WER, native diarization, word-level timestamps, smart transcription. For pure transcription, it is the dedicated tool.
How does it compare to AI tools for SEO or marketing?
It is an infrastructure component, not an end-user tool. However, teams using AI tools for SEO or marketing can use it to transcribe podcasts or webinars before repurposing them into content. The same goes for social media teams that turn live streams into posts.
✅ Conclusion
Gemini 3.5 Transcribe is the best SaaS STT model available in August 2026. The combo of a 2.6% WER in batch, native smart transcription, code-switching across 85 languages, and pricing at $0.005/minute leaves no doubt about Google's intention: to dominate speech-to-text just like it dominates search.
If you are building a voicebot, a meeting tool, or a content pipeline, test it today on AI Studio. And if you are looking for the best AI tools for code to integrate this model into your stack, we've got what you need.