Alphabet rolls out Gemini 3.5 Transcribe, aiming for more accurate real-time speech-to-text

Alphabet is introducing Gemini 3.5 Transcribe, a new speech-to-text model designed to deliver what Google describes as more precise, more intelligent, real-time transcription for voice interactions.

In a Google Blog post dated August 26, the company said the model is built to convert raw audio directly into “accurate, polished, formatted text,” with an emphasis on handling background noise, complex jargon, and disfluency cleanup that can challenge conventional speech recognition systems.

Google positioned 3.5 Transcribe as a step up from its previous transcription model, Chirp 3, describing improvements in word error rates and latency. It also cited a measured change in “time to final transcription,” saying it improves by 70% according to Artificial Analysis.

For multilingual performance, Google referenced the FLEURS benchmark, describing results across “a set of top languages and locales.” The company said the model achieves a 5.50% word error rate (WER) in streaming mode and 5.04% in non-streaming use cases, and that it performs better than Chirp 3 on the benchmark.

Beyond raw transcription quality, Google said Gemini 3.5 Transcribe is designed to preserve a user’s natural speaking style to improve intent understanding and recognize custom vocabulary, aiming to help voice interfaces carry out tasks more reliably based on what the speaker means rather than only what they say.

Google said the model is already available to consumers through its products, pointing to “the Gemini app” and Android experiences, including new voice capabilities it associated with Rambler on Android and in the Gemini app on macOS. The company framed this rollout as part of a broader effort to make voice input and inline editing feel more natural across everyday surfaces.

For developers, Google said access is available through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, and that the model is intended to fit into common workflows such as voice agents, real-time captioning tools, and post-call analytics pipelines. Google described the model as “plug[ging] seamlessly” into developer toolchains rather than requiring specialized media-handling components.

Google also described 3.5 Transcribe’s integration with Google’s Gemini Live API. It said platforms that build voice-driven interfaces can leverage the Live API so developers do not need to manage the “complex real-time media streaming infrastructure” themselves. The post named several ecosystem partners, including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents, as examples of developer platforms that can use the Gemini Live API.

Google further tied the update to in-product experience, saying context-aware understanding is brought into everyday interfaces such as Gboard, Antigravity, the Gemini app, and Chrome, including support for capturing nuances, intent, and inline edits. The company also highlighted feedback from companies it named as Vivo, Intellitek Health, and Lingopal, which it said praised 3.5 Transcribe’s latency, accuracy, and language coverage.

What Google did not spell out in the post are deployment schedules, pricing, availability timelines across regions, or detailed documentation on how “custom vocabulary” is specified in developer workflows. The model’s real-world behavior in edge cases, such as highly technical multi-speaker meetings, was also not quantified beyond the cited benchmarks and the general description of robustness to noise and disfluency.