Google AI Edge Gallery adds offline speech recognition and text-to-speech — no internet needed

Google gave its on-device AI playground a voice today: AI Edge Gallery can now listen and speak entirely offline, transcribing speech in 100+ languages and reading replies aloud without ever touching the cloud.

What we checked: Google’s own announcement, “On-Device Speech Recognition and Text-to-Speech in Google AI Edge Gallery”, dated November 10, 2026, cross-checked against LetsDataScience’s report of the same announcement.

• What: on-device Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) in Google AI Edge Gallery, fully offline.
• When: announced November 10, 2026, on the Google Developers Blog.
• Android path (ML Kit): ASR across 100+ languages at 10x real-time, 8.0% to 9.0% word error rate on FLEURS; TTS starts speaking in ~100 ms, 24 kHz audio at up to 50x real-time, 8 voices, 21 locales; 0 MB added to an APK.
• Cross-platform path (MediaPipe Tasks + LiteRT): Parakeet-TDT 0.6B for ASR; PocketTTS 100M, Kokoro-82M and Phonon 31M for TTS, the last with sub-100 ms time-to-first-audio; runs on iOS, macOS, Windows, Linux and Web.
• Showcase: XR Language Teacher points the camera at a street sign or menu and answers aloud in your language, using Gemini Nano on Android or Gemma 4 elsewhere; AI Chat gains voice dictation and read-aloud playback.
• Try it: the AI Edge Gallery app is available on Android and iOS, and the reference implementation is open source on GitHub.

Two ways to give a phone a voice#

Google ships the same offline voice pipeline two ways, and the choice is really about where your app lives. Here is how the two paths compare:

ML Kit: the Android fast lane#

For Android developers, ML Kit is the zero-effort route. Its GenAI Speech Recognition and Text-to-Speech APIs are platform-integrated: the underlying models are managed through AICore and Google’s on-device speech services, so there are no model downloads to manage and no APK size to pay for. The hardware does the heavy lifting too, with acceleration across NPU and TPU backends where available, plus background model updates.

The numbers are the story here. The recognizer streams live partial and final transcriptions as you speak — about 300 ms of work per 3 seconds of audio — and Google says it cuts end-to-end speech translation wait times by 35% to 44%. The voice side is a compact few-step diffusion model that starts speaking in roughly 100 ms and synthesizes a 27-second paragraph in about 500 ms, while cutting active device power by over 50% compared with traditional autoregressive TTS. It even ships built-in earcon support for UI sounds.

MediaPipe Tasks and LiteRT: the open-weight road#

Everything else goes through MediaPipe Tasks on top of LiteRT, Google’s on-device inference runtime. The pitch is turnkey cross-platform APIs that hide the fiddly parts — audio framing, voice activity detection, decoding — behind one interface on iOS, macOS, Windows, Linux and Web. Speech models ship as a single .litertlm file, bundling every subgraph, tokenizer and vocoder a model needs.

The bundled models are all open-weight: Parakeet-TDT 0.6B handles multilingual recognition, while PocketTTS 100M and Kokoro-82M cover multilingual speech and the 31-million-parameter Phonon handles ultra-lightweight English synthesis with sub-100 ms time-to-first-audio. And the catalog does not stop there — the Hugging Face LiteRT Community, with more than 10,000 followers, already lists Whisper, Qwen3-ASR and Moonshine for recognition and Qwen3-TTS and Matcha for synthesis. Google pairs the voice models with on-device small language models like Gemma 4 so a whole conversation can run locally.

The showcase that sells it: XR Language Teacher#

Google’s demo is a language tutor for travelers. Point the camera at an unfamiliar object or menu in a metro station or a busy market, ask aloud “What is that, and how do I ask for one?”, and the phone answers in your language with authentic pronunciation in the local one — all without a connection. On-device ASR streams the spoken question alongside the camera frame to Gemini Nano on Android, or Gemma 4 on other platforms, in a single pass, and a structured-output schema lets the TTS pipeline switch voices and accents mid-sentence between the two languages.

The same pipeline also lands in the app’s main AI Chat feature: voice dictation, read-aloud playback with live sentence highlighting, and a floating TTS bar that lets you tap any sentence to jump to it. Google’s open-source reference implementation shows how to strip Markdown, skip code snippets and tables, and schedule synthesis sentence by sentence so the voice work never fights the language model for compute. The company also points to the same setup running on smart glasses, with inference on the paired phone — a talk from Android Dev Summit walks through the adaptation.

The verdict: the phone stops needing the cloud to talk#

This is the on-device AI story maturing past chatbots. A phone that transcribes and speaks without a server is a phone that works in a tunnel, on a plane, or in a country where roaming costs a fortune — and one that never sends your voice anywhere. Android developers get the polished path today; everyone else gets the open one, with real models and real speeds rather than a research demo. Google’s frontier push kept pace the same day, with Gemini 4 Carbon reaching general availability, but the Edge Gallery release is the one that changes what a phone can do with no bars at all.

The catch is the fine print. Every speed and accuracy figure here is Google’s own, with no independent measurements published alongside it. The ML Kit path is Android-only, and the open-weight models are small by design — Phonon’s 31 million parameters buy speed, not the expressiveness of a cloud voice. And the hardware angle cuts both ways: local AI is now a selling point for devices themselves, as GIGABYTE’s 1-litre RTX Spark desktop shows. What to watch: whether developers actually ship voice-first apps on this stack, and whether the smart-glasses version becomes the first place the offline voice pipeline feels essential rather than clever.