The missing link for completely offline mobile AI agents has always been the voice layer. While mobile NPUs have grown fast enough to run 3B-to-8B text models locally, expressive voice synthesis typically required streaming audio back and forth to cloud GPUs. This week, Google’s AI Edge team solved that bottleneck.
In a release on the Google Developers Blog, the AI Edge group unveiled production-ready mobile libraries for streaming offline Automatic Speech Recognition (ASR) and lightweight diffusion Text-to-Speech (TTS) across Android (via LiteRT) and iOS.
Full-Duplex Voice AI in Airplane Mode
- Google’s AI Edge engineering team has released new open developer runtimes for offline Automatic Speech Recognition (ASR) and diffusion-based Text-to-Speech (TTS).
- Designed to pair directly with on-device Small Language Models (such as Gemma 3 and quantized Llama/Qwen variants), the pipeline enables full duplex voice conversations in airplane mode.
- Using 4-bit weight quantization and hardware NPU delegation, the entire speech-to-LLM-to-speech loop runs with under 180 milliseconds of latency.
“When speech recognition, LLM reasoning, and diffusion voice synthesis all run on the phone’s NPU, voice AI becomes free, instant, and 100% private.”—
- ✓Sources: written from primary sources and official statements, linked above.
- ✓Fact-check: names, dates and claims checked against the original sources.
- ✓AI use: researched with AI tools; written and reviewed by our editors. Our AI policy
- ✓Independence: UpdateNext has no financial relationship with the companies covered.



