Gemini 3.5 Live Translate API: Real-Time Speech Translation Enters the Routing Layer

Google shipped gemini-3.5-live-translate-preview to public preview on June 9, opening a new model endpoint inside the Gemini Live API. For teams running voice agents or multilingual conversation pipelines, the arrival changes one specific routing decision: where to send continuous speech that needs to cross a language boundary.

What happened

Gemini 3.5 Live Translate is a dedicated audio model for speech-to-speech translation. Unlike pipeline approaches that chain ASR → NMT → TTS as separate calls, this model processes an incoming audio stream and emits translated audio continuously, staying a few seconds behind the speaker throughout a session.

Key facts from the public preview launch:

• Model ID: gemini-3.5-live-translate-preview
• Access: Gemini Live API (WebSocket streaming) and Google AI Studio
• Languages: auto-detects 70+ input languages, translates across 2,000+ language pairs
• Latency posture: continuous output, not turn-based — the model does not wait for speaker to pause before emitting translated audio
• Noise robustness: documented; handles loud or unpredictable acoustic environments
• SynthID watermark: all generated audio is watermarked for AI provenance detection
• Enterprise preview: Google Meet integration in private preview starting June 2026

Developer platform partners that already expose the model via their SDK layers include Agora, LiveKit, Pipecat, and Fishjam. Each handles WebSocket streaming infrastructure so application code targets their abstraction rather than the raw Gemini Live API.

Why it matters for AI engineering teams

Until now, real-time voice translation in production meant one of two paths: a proprietary translation vendor (SYSTRAN, ModernMT, or similar) or a chained pipeline assembled from separate ASR, NMT, and TTS calls. Both paths carry integration complexity and latency compounds — every hop adds round-trip time.

gemini-3.5-live-translate-preview collapses that stack into a single streaming endpoint. The routing decision simplifies: teams that already use the Gemini API for text or multimodal workloads can add speech-to-speech translation without a separate vendor contract.

The 70+ language auto-detect eliminates the source-language configuration step that most pipeline approaches require. This matters for use cases where the speaker's language is unknown at session start — support queues, multilingual meetings, public-sector counters.

Grab's pilot is an operational signal worth watching: the company runs more than 10 million voice calls monthly for driver-to-passenger communication across Southeast Asian languages. Latency and acoustic noise are first-class constraints in that workload, not secondary concerns.

The router and operator angle

For teams using an AI gateway to manage multiple Gemini endpoints, gemini-3.5-live-translate-preview introduces a model-routing decision that did not exist before:

When to route to gemini-3.5-live-translate-preview vs. gemini-2.0-flash-live-001:

• Translation-specific sessions with known language-crossing intent → gemini-3.5-live-translate-preview is purpose-built and avoids the overhead of prompting a general model to translate
• General voice conversations with no translation requirement → gemini-2.0-flash-live-001 or equivalent remains the right path
• Mixed sessions where translation need is conditional → route on session metadata or first-utterance language detection, not model level

Streaming architecture consideration: the model operates over the Gemini Live API WebSocket path. Unlike the standard REST /v1/models/{model}:generateContent endpoint, Live API sessions are stateful and long-lived. Routing layers that proxy REST-only cannot forward these sessions without WebSocket pass-through support.

Fallback design: gemini-3.5-live-translate-preview is a preview model with no documented SLA. Production voice-translation pipelines should define a fallback: either a chained pipeline (Gemini Flash for ASR + a translation call + a TTS call) or an alternative provider. The fallback adds latency but preserves availability when the preview endpoint degrades.

Cost accounting: the Gemini Live API bills on audio input/output tokens, not request count. Continuous translation sessions that run for minutes accumulate token cost at the audio token rate. Routing layers need to surface session-duration metrics alongside per-request cost to give operators an accurate billing view.

Language pair coverage gap: the 2,000+ language pair claim covers combinations of the 70 detected languages. Not all pairs have equal quality. Teams building SLA-backed multilingual products should validate quality on their specific language pairs during the preview window rather than assuming uniform coverage.

What to watch

• GA timeline: "public preview" status means no production SLA. Google has not announced a general availability date. Monitor the Gemini API changelog.
• Google Meet rollout: the private preview for enterprise Workspace customers begins in June 2026 and could accelerate demand for the API endpoint as teams prototype integrations before enterprise access widens.
• Pricing at GA: audio token pricing for Live API during preview may change at GA. Lock in budget estimates based on preview rate card but plan for adjustment.
• Third-party SDK parity: LiveKit, Pipecat, Agora, and Fishjam have native integrations. If your voice infrastructure runs on a different WebRTC or media stack, check integration docs before assuming the model is accessible.

For teams evaluating which Gemini model to route voice-translation sessions through, the models catalog on TheRouter shows the currently routable Gemini endpoints. The Live API WebSocket path requires direct integration with Google or a platform partner that wraps it.