On June 9 Google officially launched Gemini 3.5 Live Translate, a real-time speech-to-speech translation audio model. It stems from the Gemini 3 Pro architecture, with audio input supporting 128K token context, and output delivering both text and speech simultaneously, up to 64K tokens. The most critical thing is its "translate as they speak" concurrent streaming inference method — traditional simultaneous interpretation systems must wait for the speaker to pause, then translate the whole segment before outputting, with a halting rhythm; Live Translate uses continuous stream processing, compressing latency to within a few seconds while preserving the speaker's tone, speaking speed, and pitch.
"Streaming voice translation" is not a new concept. Google Translate had demos years ago, requiring Pixel Buds paired with Android; only late last year did model-driven real-time translation extend to any earphones and iOS. But at that time, latency and tone restoration were still at the "barely usable" stage. 3.5 Live Translate's watershed is the simultaneous improvement across three evaluation dimensions: translation quality scored automatically by AutoMQM error classification; latency measured by both first-character latency and word-level latency; speech naturalness uses dedicated TTS metrics. Together they take "real-time translation" from demo to production.
The product form is worth pondering. It is not a switch hidden in Google Translate, but a standalone tunable model in the Gemini Live API — developers can build B-end scenarios like simultaneous interpretation, multilingual call transfer, and cross-border live subtitling. Google Meet's enterprise internal test rolled out before To C, followed by the full rollout of Translate on Android/iOS. This "developers first, enterprises next, then the masses" rhythm reflects Google's bet on API ecosystem monetization.
Horizontally, in May Alibaba Qwen released Qwen3.5-LiveTranslate-Flash, Tencent Hunyuan open-sourced Hy-MT2, Sakana AI released the KAME cascade architecture, and voice translation has become a standard battleground for the major model vendors. Gemini 3.5 Live Translate's differentiation is no longer "can it be done," but who can take latency, tone, and multilingual noise robustness to commercial grade at the same time. The next competition will be inference optimization for terminal chips — packing a model this large into earphones and phones is the real hard part.