--- title: "Google Ships Gemini 3.5 Live Translate Across 70 Languages in Near Real-Time" subtitle: "Google's new audio-to-audio model streams speech translation across 70+ languages in seconds, killing the transcribe-translate-synthesize pipeline for good" url: "https://alphasignal.ai/news/google-ships-gemini-3-5-live-translate-across-70-languages-in-near-real-time-2" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Google for Developers" company_url: "https://alphasignal.ai/company/googledevs" published: "2026-06-12T17:00:00.000Z" updated: "2026-06-12T22:55:18.330Z" type: "News" topics: - "Api" - "Audio" subtopics: - "Realtime Voice" - "Text To Speech" - "Speech To Text" upvotes: 172 read_time: "6 min" source_url: "https://accounts.google.com/v3/signin/identifier?continue=https%3A%2F%2Faistudio.google.com%2Flive%3Fmodel%3Dgemini-3.5-live-translate-preview%26linkId%3D62288739&dsh=S-821561662%3A1781304804069119&followup=https%3A%2F%2Faistudio.google.com%2Flive%3Fmodel%3Dgemini-3.5-live-translate-preview%26linkId%3D62288739&passive=1209600&flowName=WebLiteSignIn&flowEntry=ServiceLogin&ifkv=AcDsRvxO43lb9cKKo0rxggmeDYyz09cUw53_Y7YzqYnO5yaNbdHDanvA_GPULzsXCAPbsb_rChiM" comments: 0 --- # Google Ships Gemini 3.5 Live Translate Across 70 Languages in Near Real-Time > Google's new audio-to-audio model streams speech translation across 70+ languages in seconds, killing the transcribe-translate-synthesize pipeline for good Google just shipped [Gemini 3.5 Live Translate](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-live-3-5-translate/), a dedicated speech-to-speech model that translates spoken audio in near real-time across more than 70 languages. It is now in public preview via the Gemini Live API and Google AI Studio, and it represents the most significant architectural shift in machine translation in years. ## The pipeline is dead. Long live the pipeline. Every real-time translation system you have ever built or used almost certainly followed the same three-step chain: speech-to-text (ASR), text translation (MT), then text-to-speech (TTS). Each hop adds latency, loses acoustic nuance, and introduces a new failure mode. Google dropped that cascaded pipeline entirely in favor of a single audio-to-audio model. Gemini 3.5 Live Translate is a low-latency, audio-to-audio model optimized for real-time translation of spoken conversations, enabling seamless bidirectional translation with high accuracy and natural voice output. The model ID is `gemini-3.5-live-translate-preview`, and it is audio-to-audio rather than the older speech-to-text-to-translation-to-text-to-speech pipeline, accepting streamed source audio in 100-millisecond chunks and producing translated speech as output. The practical difference is not just latency. The translated voice sounds closer to how the original speaker actually sounds, rather than flat synthesized audio -- preserving intonation, tempo, and vocal pitch. In a meeting with multiple participants, listeners can still distinguish individual voices and register the emotional register of what is being said. ## What it can do - Turn-by-turn systems wait for the speaker to finish before translating, which adds latency. Gemini 3.5 Live Translate processes speech as it streams and generates translated audio continuously, staying just a few seconds behind the speaker without awkward pauses. - The model handles multilingual inputs without the need to manually configure settings. Source language is detected automatically. - Noise robustness ensures applications can handle loud, unpredictable environments. - Every audio output is watermarked using Google's SynthID technology, embedding an invisible marker in the audio to flag it as AI-generated. - Optional transcripts for both input and output are available as a sidecar alongside the spoken audio. ## What it cannot do The model card is refreshingly candid about limitations. According to the [official Google DeepMind model card](https://deepmind.google/models/model-cards/gemini-3-5-audio/), voices can be inconsistent: they may shift after long pauses, change gender, or get stuck on one voice during rapid multi-speaker sessions. Language detection can struggle with non-native accents, similar languages, or rapid language switches. There are also hard API constraints worth knowing before you build: - To hold strict real-time latency thresholds, the translation path accepts audio input only -- text input is not supported in translation mode. - The model drops tool use and system instructions in this mode, keeping it a focused translator pipeline rather than a general agent. - It is a preview release: Google has not committed to production-grade SLAs, stable pricing, or general availability. ## Under the hood The model is built on Gemini 3 Pro , fine-tuned for the specific constraints of real-time audio streaming. Gemini 3.5 Live Translate was evaluated across three main quality dimensions: translation quality, latency, and speech naturalness. Translation quality uses AutoMQM -- an error-based automatic metric which identifies and categorizes translation errors to produce a fine-grained quality score. Latency is measured at two granularities: initial latency (time from first spoken word to first translated word) and word-level latency (average delay per word through the session). It is a dedicated Live API translation mode: stream 16kHz PCM speech in, choose a target language, and receive translated 24kHz audio plus optional transcripts from `gemini-3.5-live-translate-preview`. Here is the minimal setup in JavaScript: ``` const config = { responseModalities: [Modality.AUDIO], inputAudioTranscription: {}, outputAudioTranscription: {}, translationConfig: { targetLanguageCode: 'pl', echoTargetLanguage: true } }; const session = await ai.live.connect({ model: 'gemini-3.5-live-translate-preview', config: config, callbacks: { onmessage: (message) => { const content = message.serverContent; // Translated audio arrives as base64-encoded chunks const audioChunk = content?.modelTurn?.parts[0]?.inlineData?.data; } } }); ``` ## Where it is available right now Gemini 3.5 Live Translate is rolling out across Google products: for developers in public preview via the Gemini Live API and Google AI Studio, for enterprises in private preview starting this month in Google Meet, and for everyone via Google Translate on Android and iOS. The Google Meet upgrade is the most structurally significant change for enterprise users. In Google Meet specifically, it enables 70+ languages and over 2,000 language combinations within a single meeting -- up from five languages that could only translate to and from English. Google Meet users on Workspace plans get a private preview starting this month, with a full rollout later in the year. ## The ecosystem play Google is not asking you to build the media infrastructure yourself. Agora, Fishjam, LiveKit, Pipecat, and Vision Agents have all connected to the Gemini Live API, allowing developers building voice translation applications to use these platforms' infrastructure for handling real-time media streams. This separates the media engineering problem from the translation problem: platform providers handle streaming complexity while developers focus on the user experience layer. The real-world deployments are already compelling. Grab is testing the model to enable multilingual communication in near real-time between drivers and travelers at pickups. These users make over 10 million voice calls per month through Grab. That is a concrete scale test for a capability that previously required human interpreters or was simply unavailable. ## Pricing and access API pricing is $0.023 per minute, which undercuts several competing services. You can [try it directly in Google AI Studio](https://aistudio.google.com/live?model=gemini-3.5-live-translate-preview) without writing a line of code, or clone the [Gemini Cookbook](https://github.com/google-gemini/cookbook) for working example apps. For client-side applications, you must use the v1alpha endpoint with ephemeral tokens, and you should specify the `translationConfig` in the token creation constraints on your server to ensure the translation configuration is locked and cannot be tampered with by the client. ## What this means for the field The assumption that real-time translation requires a multi-model stack is now obsolete. A single model, a WebSocket connection, and 16kHz PCM audio is all you need to ship a live translation feature. The latency, voice fidelity, and noise robustness that used to require significant infrastructure investment are now a two-line config change away. The remaining question is not whether this architecture works -- it clearly does -- but how quickly the preview becomes a production-grade API that teams can actually depend on. ## Sources - [X post by @googledevs](https://x.com/googledevs/status/2065479241215488401) - [Fluid, natural voice translation with Gemini 3.5 Live Translate – Google Blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-live-3-5-translate/) - [Gemini 3.5 Audio (Live Translate) – Model Card, Google DeepMind](https://deepmind.google/models/model-cards/gemini-3-5-audio/) - [Live translation with Gemini Live API – Google AI for Developers](https://ai.google.dev/gemini-api/docs/live-api/live-translate) - [Google's Gemini 3.5 Live Translate lands with 70-language real-time speech – PPC Land](https://ppc.land/googles-gemini-3-5-live-translate-lands-with-70-language-real-time-speech/) - [Gemini 3.5 Live Translate: Real-Time Multilingual CX – DigitalApplied](https://www.digitalapplied.com/blog/gemini-live-3-5-translate-real-time-multilingual-cx-guide)