Gemini 3.5 Transcribe Hits 2.6% WER, Cuts Latency 70% vs Chirp 3
In this article
Google DeepMind shipped Gemini 3.5 Transcribe on August 26, 2026, as a purpose-built replacement for conventional automatic speech recognition in developer pipelines. The model is in public preview through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. For teams already thinking about how pipeline architecture shapes AI output quality, a transcription layer that handles disfluency cleanup, jargon adaptation, and multi-language detection natively removes a category of preprocessing work that typically sits upstream of every downstream task.
Two APIs, Two Latency Profiles
The model ships with a hard architectural split between its two delivery modes. Real-time audio routes through the Live API under the model identifier gemini-3.5-transcribe-live, providing continuous bidirectional streaming at sub-second latency — the profile appropriate for voice agents and live captioning. Pre-recorded audio uses the Interactions API under gemini-3.5-transcribe, adding speaker attribution and word-level timestamps for post-call analytics and meeting transcription. This is not a single endpoint with a latency toggle; the two modes carry different performance characteristics and developers choose based on pipeline requirements.
Benchmark Numbers
As measured by Artificial Analysis, Gemini 3.5 Transcribe achieves a Word Error Rate of 4.0% for streaming and 2.6% for non-streaming. On the FLEURS multilingual benchmark across a set of top languages and locales, the model scores 5.50% WER in streaming mode and 5.04% WER in non-streaming mode. Time to final transcription improved by 70% over Chirp 3, also as measured by Artificial Analysis. Google notes strong accuracy on alphanumeric entities — postal codes, order IDs — historically a weak point for general-purpose ASR. Custom vocabulary injection allows domain-specific jargon and non-standard spellings to be passed at inference time without fine-tuning.
| Model | Streaming WER | Non-streaming WER | FLEURS Streaming WER | FLEURS Non-streaming WER | Time-to-final vs. Chirp 3 | Language Coverage | Speaker Attribution |
|---|---|---|---|---|---|---|---|
| Gemini 3.5 Transcribe | 4.0% | 2.6% | 5.50% | 5.04% | −70% | 85+ languages | Up to 3 speakers (3+ experimental) |
| Chirp 3 (predecessor) | Not disclosed | Not disclosed | Higher than 5.50% | Higher than 5.04% | Baseline | Not disclosed | Not disclosed |
Smart Transcription and Pipeline Impact
Gemini 3.5 Transcribe applies a layer Google calls "smart transcription": automatic removal of filler words such as "um" and "ah," resolution of self-corrections (handling phrasing like "let's meet Tuesday—no, Wednesday" by retaining only the intended content), and auto-formatting of output text. For developers building downstream NLP or summarization stages, text arriving at the next pipeline step is already cleaned, potentially changing prompt engineering requirements at subsequent stages.
The model also supports function calling — delegating tasks such as image generation or file analysis to other Gemini models — though this capability is currently restricted to the Gemini macOS app. Automatic language detection operates across 85-plus languages, handling regional accents and mid-conversation language switches without manual language tagging in the request. Multi-speaker identification in pre-recorded audio supports up to three speakers with word-level timestamps; support for more than three speakers is marked experimental.
Platforms including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents have integrated the Live API ahead of launch, managing real-time media infrastructure so that application developers interact only with transcription output.
Availability
Developer access is through the Gemini API in Google AI Studio and Google Antigravity, both in public preview. Enterprise access is through the Gemini Enterprise Agent Platform, also public preview, with a forthcoming rollout to Gemini Enterprise for Customer Experience. Consumer surfaces include Rambler on Android — which uses 3.5 Transcribe to convert spoken input into formatted text and accept inline voice edits — the Gemini macOS app in English, and a Chrome integration for dictation into arbitrary web fields listed as coming soon.
The 70% latency reduction over Chirp 3 and the dual-API architecture indicate Google is treating transcription as a core infrastructure component. As voice interfaces move toward the default interaction layer in agentic systems — a trajectory consistent with what the agentic era demands from data pipelines — a transcription model with native jargon adaptation, disfluency handling, and function-calling hooks becomes a first-class pipeline component rather than a peripheral utility. The FLEURS and Artificial Analysis figures are third-party measurements, but they are the only quantitative comparisons Google has published against Chirp 3.