Google DeepMind Launches Gemini 3.5 Transcribe Speech-to-Text Model

Google DeepMind on Aug. 26 released Gemini 3.5 Transcribe, a speech-to-text model that the company says goes well beyond raw audio transcription. Rather than treating spoken words as a signal-processing problem, the model applies reasoning to resolve self-corrections, strip filler words, format output, attribute multiple speakers and even delegate follow-on tasks — like image generation or file analysis — to other Gemini models via function calls.

The announcement replaces Chirp 3 as Google’s flagship transcription offering and, notably, consolidates speech-to-text onto the same Gemini API surface that developers already use for text and multimodal tasks. That means no separate SDK to learn: the model is callable with the identifiers gemini-3.5-transcribe for batch audio and gemini-3.5-transcribe-live for real-time streaming, and audio is billed in tokens like any other Gemini input.

What the Benchmarks Actually Say

According to measurements by third-party benchmarking firm Artificial Analysis, Gemini 3.5 Transcribe achieves a word error rate of 4.0% in streaming mode and 2.6% in non-streaming mode. On the FLEURS multilingual benchmark — a standard test across a broad range of languages and regional accents — the model posts 5.50% WER streaming and 5.04% WER non-streaming. Time to final transcription improves by 70% over Chirp 3 on the same benchmark suite.

Those numbers deserve some context. Word error rate treats every missed word equally, meaning a skipped filler word counts the same as a missed medication name. Cross-vendor comparisons should be read carefully for that reason. Still, the trajectory is clear: Google has substantially closed the gap with specialist providers while adding capabilities — speaker diarization, word-level timestamps, custom vocabulary and function calling — that most dedicated transcription APIs charge extra for or don’t offer at all.

A Crowded Market With Real Competition

The speech-to-text space moved fast before this announcement. Microsoft shipped MAI-Transcribe-1 in April 2026, reporting a 3.8% average WER on FLEURS across 25 languages, available through Azure AI Foundry. OpenAI launched GPT-Realtime-Whisper in May 2026, a streaming-optimized model priced at $0.017 per minute. Specialist provider AssemblyAI’s Universal-3.5 Pro posts a 7.69% average WER on a normalized code-switching benchmark, outpacing ElevenLabs Scribe v2 at 8.77% and Deepgram Nova-3 at 12.22%.

What separates Gemini 3.5 Transcribe from those rivals is less about raw WER and more about ecosystem depth. Competitors like AssemblyAI offer richer speech intelligence add-ons — sentiment analysis, PII redaction, entity detection — and an LLM gateway that pipes transcripts into leading language models without additional integration work. Google’s counter-argument is integration breadth: the model already powers Rambler on Android, the Gemini app on macOS, and Google Antigravity, with Chrome dictation support announced as coming soon. For developers already building in the Google stack, that surface area matters.

Why Students Should Pay Attention

The most immediately relevant detail for students and career changers is the free tier. Gemini 3.5 Transcribe is available in public preview through Google AI Studio at no cost, with a Build mode that lets developers vibe-code voice applications directly in the browser. That’s a zero-dollar entry point for portfolio projects that would otherwise require paid API credits.

The two-API structure — real-time streaming versus batch processing — also mirrors exactly the architecture patterns covered in AI and machine learning courses. Building with both modes in a single project demonstrates an understanding of latency-accuracy tradeoffs, which is the kind of concrete, explainable skill that reads well in technical interviews.

Practical project ideas enabled by the free tier include: live captioning tools for accessibility portfolios, lecture-to-structured-notes pipelines, multilingual interview or language-learning apps (the model supports over 85 languages and auto-detects them), and post-meeting analytics systems with speaker attribution. The model’s integrations with developer frameworks including LangChain, LiveKit, Pipecat and Vercel — all named in Google’s announcement — mean students can plug Gemini 3.5 Transcribe into tools they’re likely already learning without starting from scratch.

For those exploring multi-model orchestration specifically, the function-calling capability is worth experimenting with. The transcription model can hand off tasks to other Gemini models in the background, which means a single voice command can trigger a chain of AI actions — the kind of agentic workflow pattern that is increasingly common in production AI systems and worth understanding before you’re asked about it in a job interview.

The Bottom Line

Gemini 3.5 Transcribe is a meaningful step forward in speech recognition, but its real significance for the developer community is structural: transcription now lives on the same API, the same pricing model and the same reasoning infrastructure as the rest of the Gemini ecosystem. For students, that consolidation lowers the barrier to building serious voice-enabled applications to the cost of a Google account. The free tier in AI Studio is the place to start.

Source: Google DeepMind