Today, we are introducing on-device Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) capabilities in Google AI Edge Gallery. Voice is often the most natural way to interact with AI on phones and wearables. Recent advances in compact speech architectures now make it possible to run accurate, multilingual ASR and expressive TTS models entirely on device, all within a small memory footprint, with low latency, and completely offline. By pairing these lightweight speech models with on-device small language models (SLMs) like Gemini Nano and Gemma 4, developers can build apps that stream live transcriptions as the user speaks and synthesize natural spoken replies locally.
This release brings together major advancements across both Google’s system speech models and the open-weight ecosystem:
Below, we’ll explore how this local voice pipeline powers the new XR (Extended Reality) Language Teacher showcase, upgrades AI Chat in Google AI Edge Gallery, and how you can build with these APIs and models today.
Learning a language while traveling happens in the real world, often in underground stations, busy markets, or abroad where Wi-Fi and roaming data may not be available. To show how vision, language, and speech models can work together 100% locally on your phone with no network connection needed, we built the XR Language Teacher showcase in Google AI Edge Gallery. Simply point your camera at an unfamiliar object or menu, ask "What is that, and how do I ask for one?", and immediately hear an explanation in your language alongside authentic pronunciation in the local language.
Link to Youtube Video (visible only when JS is disabled)
With inference running locally on a paired phone, the same experience works standalone on a phone and extends naturally to hands-free smart glasses. To see how we adapted Language Teacher for intelligent glasses check out our recent talk at Android Dev Summit: Building Audio-First Experiences for Intelligent Eyewear with Gemini Nano.
Beyond the XR Language Teacher showcase, we also built ASR and TTS directly into Google AI Edge Gallery’s main AI Chat feature, giving on-device SLMs like Gemini Nano and Gemma 4 a natural, offline voice interface while on the go.
Link to Youtube Video (visible only when JS is disabled)
While multimodal models like Gemma 4 can process audio inputs directly, we provide a modular solution for cascading language models with separate, lightweight voice models for latency optimization and flexibility. Depending on your target devices and model requirements, our open-source code on GitHub shows you the optimized and recommended solution for Android using ML Kit, as well as MediaPipe Tasks and LiteRT for cross-platform coverage and open model support.
For developers building for Android devices, ML Kit’s GenAI Speech Recognition and Text-to-Speech APIs provide a standardized, platform-integrated path out of the box. With the underlying speech models managed through AICore and Google’s on-device speech services, developers get production-grade quality without having to manage model downloads or worry about APK size bloat. Your app also benefits automatically from hardware acceleration across NPU and TPU backends (when available) and seamless background model updates.
Integrating ASR and TTS into cross-platform applications usually requires handling low-level audio framing, Voice Activity Detection (VAD), and complex decoding pipelines. MediaPipe Task simplifies this by providing unified, turnkey ASR and TTS APIs across iOS, macOS, Windows, Linux, and Web that abstract away manual pre-processing, audio stream routing, and platform-specific audio management into simple, high-level interfaces:
Under the hood, ML Kit and MediaPipe Tasks are powered by LiteRT, Google’s high-performance on-device AI inference infrastructure. Cross-platform applications requiring even more fine-grained control over speech processing or custom model architectures can also build on it directly.
Speech models often rely on multiple subgraphs and auxiliary assets, such as encoder/decoder subgraphs and tokenizers for ASR, or phonemizer dictionaries, voice embeddings, and vocoders for TTS. All required components for each model are packaged into a single .litertlm file for easy deployment. The app ships with ready-to-use .litertlm models, including Parakeet-TDT-0.6B-v3 (multilingual ASR), PocketTTS 100M (multilingual TTS), Phonon 31M (ultra-lightweight English TTS), and Kokoro-82M (multilingual TTS). Beyond these, the vibrant Hugging Face LiteRT Community with over 10K+ followers offers a growing catalog of open-weight models, like Whisper, Qwen3-ASR and Moonshine for ASR, or Qwen3-TTS and Matcha for TTS. For developers seeking custom implementations or support for additional open-weight voice models, LiteRT provides a rich set of multimodal model examples to help you get started, with more on the way.
Adding a local voice pipeline takes on-device SLMs beyond typing-based chat interfaces, enabling natural, voice-first interactions that work anywhere, completely offline. All the building blocks to try these experiences or integrate them into your own apps are available today:
We'd like to extend a special thanks to our significant contributors for their work on this project:
Alice Zheng, Ajay Dudani, Alex Kanaukou, Amy Zeppenfeld, Byungchul Kim, Cheng Yang, Chintan Parikh, Chris Parsons, Cormac Brick, David Chou, Dong Chen, Geonsun Lee, Glenn Cameron, Hriday Chhabria, Ian Ballantyne, Ireneu Pla, James O’Donnell, Jamieson Pryor, Jason Poon, Jenn Lee, Jianing Wei, Jing Jin, Jingxiao Zheng, Jingyu Shi, Karthik Thirumalai, Kristen Wright, Lu Wang, Marius Kintel, Matthias Grundmann, Miao Wang, Mogan Shieh, Na Li, Paul Hsia, Qiumin Xu, Ravi Shankar, Rishika Sinha, Ronald Wotzlaw, Sachin Kotwani, Sandeep Patil, Sebastian Schmidt, Shao-Fu Shih, Shuangfeng Li, Suril Shah, Tiger Souvannakoumane, Tzu-Ping Chen, Weiyi Wang, Wentao Tu, Yang Yang, Yi-Chun Kuo, Yishuang Pang, Yu-che Liu, Yu-hui Chen, Yuling Liang, Zhihao Yang, Zichuan Wei.
