Editorial illustration for Google AI Claims 2.6% Average WER for Gemini 3.5 Transcribe
Google's Gemini 3.5 Transcribe Hits 2.6% Error Rate
Google AI Claims 2.6% Average WER for Gemini 3.5 Transcribe
Google put a number on its new transcription model, and it's a specific one: 2.6% average word error rate, measured by Artificial Analysis, across more than 85 languages. The model, Gemini 3.5 Transcribe, arrived as two separate endpoints rather than a single product. `gemini-3.5-transcribe` handles pre-recorded audio through the Interactions API, while `gemini-3.5-transcribe-live` handles bidirectional streaming through the Live API.
Google says streaming comes in at 4.0% WER, and that time to final transcription is 70% faster than Chirp 3, the model it replaces. The system also detects language automatically, including switches mid-sentence, which matters for call centers and multilingual meetings.
None of this ships as open weights. There's no self-hosted option, no checkpoint to download. Access runs entirely through Google's API, meaning the decision facing developers isn't about infrastructure, it's about which endpoint fits the job and which pricing tier covers it. The two endpoints don't share limits or feature sets, and that split is the detail worth understanding before building anything on top of it.
Google reports average word error rates of 4.0% streaming and 2.6% non-streaming, as measured by Artificial Analysis. Time to final transcription improves 70% over Chirp 3, the previous model. Automatic detection covers more than 85 languages, including mid-sentence code-switching.
Why this matters
Splitting transcription into two endpoints, one for files, one for live audio, tells us Google is optimizing for production use cases rather than benchmark bragging rights. The 2.6% average WER across 85+ languages is the headline number, but the streaming figure, 4.0% per Artificial Analysis, is the one builders actually need to plan around. Real-time voice interfaces live or die on that gap between batch and live accuracy, and a 1.4-point spread is not trivial once you're routing customer calls or transcribing meetings as they happen.
The line about readable summaries and auditable transcripts now requiring separate API calls is worth sitting with. It means Google is drawing a formal boundary between "what a human wants to read" and "what a compliance team needs to verify," which is a design choice, not a limitation. For developers building anything regulated, that's a decision about liability as much as UX. We'd want independent verification of these WER numbers before treating them as ceiling performance, especially across the lower-resource languages buried inside that 85+ count.
Related Reading
Browse all Google news →
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI
From Archives Vercel AI's vgpu Open-Sources WGSL Module System for TypeScript WebGPU
From Archives OpenAI's Codex Gains a "Persistent Mode" in New Code
From Archives Google’s AI Notebook app adds full-text access to over 100,000 books
LLMs & Generative AI
From Archives Barret Zoph Leaves OpenAI for Google After Five Months
Business & Startups
Common Questions Answered
What is the word error rate (WER) difference between Gemini 3.5 Transcribe's streaming and non-streaming modes?
Gemini 3.5 Transcribe achieves a 2.6% average WER for pre-recorded audio through the non-streaming endpoint, while the streaming endpoint reports a 4.0% WER according to Artificial Analysis measurements. This 1.4-point difference between batch and live accuracy is significant for builders planning real-time voice interfaces.
How many languages does Gemini 3.5 Transcribe support, and what unique capability does it include?
Gemini 3.5 Transcribe provides automatic language detection across more than 85 languages, including the ability to handle mid-sentence code-switching. This multilingual support represents a substantial expansion in transcription capabilities across diverse language pairs and mixed-language scenarios.
What are the two separate endpoints for Gemini 3.5 Transcribe and what is their intended use?
The `gemini-3.5-transcribe` endpoint handles pre-recorded audio files through the Interactions API, while the `gemini-3.5-transcribe-live` endpoint manages bidirectional streaming through the Live API. This split design allows Google to optimize each endpoint for its specific production use case rather than creating a one-size-fits-all solution.
How much faster is Gemini 3.5 Transcribe compared to the previous Chirp 3 model?
Google reports that time to final transcription improves by 70% with Gemini 3.5 Transcribe over the previous Chirp 3 model. This significant performance improvement makes the new model substantially more efficient for both batch and real-time transcription workflows.
Further Reading
- Papers with Code - Latest NLP Research - Papers with Code
- Hugging Face Daily Papers - Hugging Face
- ArXiv CS.CL (Computation and Language) - ArXiv
