Google's Gemini 3.5 Live Translate Is Impressive. Here's the Production Problem It Doesn't Solve.
On June 9, 2026, Google launched Gemini 3.5 Live Translate, a real-time speech-to-speech translation model covering 70+ languages. It streams translated audio continuously, preserving the speaker's intonation, pacing, and pitch. Partners like Grab, which handles over 10 million voice calls per month, are already testing it in production.
It is a genuinely impressive piece of AI engineering. It is also a perfect case study for the production problem the AI voice industry keeps ignoring.
The Industry Conflates Model Quality With Production Quality
Choosing the best model at generation time is not the same as shipping high-quality audio. The gap between those two things is where production voice AI actually fails: mispronounced product names, voice inconsistency at clip 51, multilingual artifacts, provider outages with no retry logic.
None of these are model problems. They are pipeline problems. A production-grade voice pipeline does five things a generation pipeline does not: pre-generation planning, multi-model routing, post-generation validation, automated retry, and delivery-ready output.
This is what Onepin does. It is the orchestration and validation layer that sits on top of 100+ TTS models worldwide, including Google Cloud TTS, ElevenLabs, Deepgram Aura, Cartesia, and others. If you produce voice content at scale, build the pipeline first. Start at onepin.ai.
Frequently asked questions
- What is Gemini 3.5 Live Translate?
- Gemini 3.5 Live Translate is a real-time speech-to-speech translation model Google launched on June 9, 2026, covering 70+ languages. It streams translated audio continuously while preserving the speaker's intonation, pacing, and pitch. Partners like Grab, which handles over 10 million voice calls per month, are already testing it in production.
- Why does choosing the best model not guarantee high-quality audio?
- Choosing the best model at generation time is not the same as shipping high-quality audio. Production voice AI still fails on mispronounced product names, voice inconsistency partway through a batch, multilingual artifacts, and provider outages with no retry logic. These are pipeline problems, not model problems.
- What does a production-grade voice pipeline do that a generation pipeline does not?
- A production-grade voice pipeline does five things a generation pipeline does not: pre-generation planning, multi-model routing, post-generation validation, automated retry, and delivery-ready output. Together these close the gap between a model that scores well and audio that is safe to ship.
- How does Onepin fit alongside a model like Gemini 3.5 Live Translate?
- Onepin is the orchestration and validation layer that sits on top of 100+ TTS models worldwide, including Google Cloud TTS, ElevenLabs, Deepgram Aura, and Cartesia. Rather than replacing a model, it adds the planning, routing, validation, and retry steps that make voice output production-safe at scale.