Google AI Edge Gallery adds fully offline speech recognition and text-to-speech

What it's about

On November 10, 2026, Google added two features to its Google AI Edge Gallery app that you'd expect from an AI app but that were missing in this form until now: speech recognition (ASR, Automatic Speech Recognition, which turns spoken language into text) and speech output (TTS, text-to-speech, which reads text aloud). The key point: both run entirely on your phone, without any connection to a server. So you can talk to a local AI model like Gemini Nano or Gemma 4 and get a spoken answer, even in airplane mode, on the subway or abroad without roaming.

AI Edge Gallery is an open-source Google app (Apache 2.0 license) that lets you try out small language models directly on your phone. It's available for Android 12 and later and iOS 17 and later, via the Play Store and App Store. According to the French blog Korben, the app had more than 500,000 downloads on GitHub just two months after launch.

Beyond the app update, the announcement is aimed mainly at developers: Google is providing matching APIs for Android as well as iOS, macOS, Windows, Linux and the web.

Two approaches: system models on Android, open models everywhere else

Google is taking two approaches.

On Android, Google's own speech models do the work. They're built into the operating system and accessed through ML Kit (Google's toolkit for machine learning in apps). Because the models are part of the system and managed by the AICore service, the app itself doesn't get any bigger. Google talks about "0 MB" of added APK size. Model updates arrive automatically in the background, and where available, the models use dedicated AI chips (NPU or TPU) for acceleration.

For all other platforms, Google relies on freely available speech models packaged in its own .litertlm file format. Each file contains everything a speech model needs. For speech recognition, that's the encoder, decoder and tokenizer (which breaks text into small pieces). For speech output, it's pronunciation dictionaries, voice profiles and the vocoder, which generates the audible signal at the end. You access these models through MediaPipe Tasks, a cross-platform API that automatically handles annoying details like splitting the audio stream into chunks or voice activity detection (VAD, where the system notices when someone is speaking and when there's a pause). Underneath it all sits LiteRT, Google's lightweight runtime for on-device AI.

The numbers, according to Google

All of the following figures come from Google:

The word error rate tells you how many words get recognized incorrectly. FLEURS is a widely used multilingual test dataset. Android's speech output is based on a few-step diffusion model: instead of generating sound piece by piece in sequence (autoregressively), it "computes" the audio out of noise in just a few passes. According to Google, that's where the speed and power savings come from.

Google also points to the LiteRT community on Hugging Face, which has more than 10,000 followers and offers additional models in the right format, including Whisper, Qwen3-ASR and Moonshine (speech recognition) as well as Qwen3-TTS and Matcha (speech output).

What the app can do now

XR Language Teacher: This new showcase project is meant to demonstrate how vision, speech and language models work together completely offline. You point the camera at a street sign, a menu or an object and ask something like: "What is this, and how do I order one?" Speech recognition sends the question along with the camera image to Gemini Nano (Android) or Gemma 4 (other platforms). The model answers using a fixed output schema: the explanation comes in your own language, and foreign-language expressions are marked separately. That lets the speech output switch voice and accent mid-sentence and pronounce the word correctly in the local language, while a vocabulary card pops up on screen. According to Google, it also works with smart glasses, with the processing happening on the paired phone.

AI Chat with voice: The regular chat now has voice input and output too. In model management, you pick the speech recognition, speech output, voices and language packs. With "Auto-play TTS" or a speaker button, replies get read aloud. The sentence currently being spoken is highlighted, and you can tap to jump to a sentence. The open-source code also shows how to strip Markdown formatting, skip code blocks and tables when reading aloud, and schedule speech output sentence by sentence so it doesn't compete with the language model for compute and memory.

For context on the app's other features (according to Korben and Caschys Blog): there's already Ask Image (image analysis), Audio Scribe (transcription and translation of recordings), a Prompt Lab with performance benchmarking, a Thinking mode for Gemma 4 and extensible "Agent Skills", for things like Wikipedia lookups or calendar entries, plus experimental support for MCP (Model Context Protocol, a standard for connecting external tools).

Why separate speech models?

Models like Gemma 4 can actually process audio directly. Google still opts for a pipeline of separate, small speech models plus the actual language model. The reasoning: lower latency and more flexibility. You can swap speech recognition and speech output depending on the device, and you hear the first words while the language model is still writing.

A real-world test from Japan

An author on note.com tested how well local AI on a phone holds up in everyday work, in airplane mode. He used the existing Audio Scribe and AI Chat features, though, not the new speech features. He had 20 deliberately patchy voice memos about workplace improvements turned into meeting documents. The result: the model could summarize topics and spot connections, such as between late material deliveries and process delays. But even after refining the instructions, problems remained: individual points got lost under broader topics, the model invented background details ("the management system has become a mere formality"), tables had the wrong number of columns, the Japanese text was clunky in places, and the output cut off midway. The author also stresses that working in airplane mode doesn't prove the app never sends data.

Our take

With this update, Google closes an important gap: local AI on phones has mostly meant typing until now. With voice input and output that work offline and with low latency, apps like travel language helpers or hands-free assistants on glasses become realistic, with a privacy bonus, since conversations never leave the device. It's notable that besides its own system models, Google also supports third-party open models such as Parakeet (originally from Nvidia) and Kokoro, and has open-sourced the code.

The caveats: the most impressive figures apply to Android with Google's system models and come from Google itself, with no independent verification yet. How well speech recognition works in, say, German or in noisy environments remains to be seen. And the real-world test shows that the small local language models behind the voice are significantly weaker than large cloud models, so you still need to check their answers critically.