Your phone can now listen to you and talk back without ever touching the internet. Not a demo. Not a research paper. Google shipped it today — on-device speech recognition and text-to-speech in Google AI Edge Gallery, announced November 10, 2026 on the Google Developers Blog.

I have not run any of this hands-on yet — it landed hours ago — and this article is built from Google's announcement, checked November 10, 2026. Where a number is Google's, I'll say so, because a launch-day benchmark is a vendor claim, not an independent result.

What actually shipped

Two paths, and the split matters more than the headline.

On Android, the system does the work. ML Kit's new GenAI Speech Recognition and Text-to-Speech APIs run on models Google manages through AICore, built into the OS. Google says that adds 0 MB to your APK — no model downloads to manage, no size bloat to explain to your users, and the models update in the background. Hardware acceleration across NPU and TPU backends where the device has them.

Everywhere else, the models are open-weight. MediaPipe Tasks and LiteRT ship .litertlm bundles — single files containing every subgraph, tokenizer, and voice embedding a model needs. Parakeet-TDT 0.6B for speech recognition; PocketTTS 100M, Kokoro-82M, and Phonon 31M for speech synthesis. They run on iOS, macOS, Windows, Linux, and Web. The Hugging Face LiteRT Community also lists Whisper, Qwen3-ASR, Moonshine, Qwen3-TTS, and Matcha if you want to bring your own.

Both paths pair with on-device small language models — Gemini Nano on Android, Gemma 4 across platforms — so the full loop (hear → understand → answer out loud) can run with no network at all.

The numbers Google is leading with

The speech recognition side: live streaming transcription across 100+ languages, at 10x real-time speed — about 300 ms to transcribe 3 seconds of speech — with 8.0% to 9.0% word error rate on FLEURS, per Google. Google also says it cuts end-to-end speech translation wait times by 35% to 44%.

The speech synthesis side: a compact few-step diffusion model that starts speaking in about 100 ms and generates 24 kHz audio at up to 50x real-time speed. A 27-second paragraph synthesizes in roughly 500 ms. Eight expressive voices, 21 locales, and Google claims it cuts active device power by over 50% compared to traditional autoregressive TTS. The ultra-lightweight Phonon 31M model gets sub-100 ms time-to-first-audio on most devices.

Read those numbers the way you should read any launch-day claim. FLEURS is a real benchmark, but the scores are Google's, on hardware Google chose, and nobody independent has reproduced them yet. The 100 ms figure is time-to-first-audio, not the whole pipeline — your app still has to get the words to the model first. What makes the numbers interesting anyway is which ones Google chose to lead with: latency and power, not just accuracy. That's the shape of a feature built for real phones, not a leaderboard.

Why this is different from every cloud voice API

Here's the part that actually changes what you can build. A cloud voice API means a round trip for every word: microphone → network → server → network → speaker. It means latency you can't control, a bill per minute, and audio leaving the user's device — which is a non-starter for health apps, kids' apps, and anything handling sensitive conversations.

On-device flips all three. No round trip, so the latency floor is the model, not the network. No per-minute bill. And the audio never leaves the phone, which is the kind of privacy claim you can actually defend in a review instead of just asserting in one.

Google's own showcase makes the point well. The XR Language Teacher in AI Edge Gallery points your camera at a street sign or a menu, you ask out loud what it says and how to ask for one, and you hear the answer in your language plus the pronunciation in the local one — vision, language, and speech, 100% locally. It works in an underground station, in a market abroad, anywhere the Wi-Fi doesn't. The same pipeline is now in AI Chat inside the Edge Gallery app, with sentence-synchronized read-aloud playback and a floating TTS bar.

What I'd actually build with this

If I were picking a weekend project off this announcement, it would be one of these:

  • A voice-first field app. Think warehouse checklists, inspection notes, or a language tutor for travelers — anywhere the user is standing somewhere with bad signal and both hands busy. The offline part isn't a bonus here, it's the whole point.
  • A private voice journal or therapy-adjacent app. The kind of app where "your audio never leaves the device" is the feature users actually pay for. On-device TTS means the app can also read entries back without a server in the middle.
  • A kids' reading tutor. Streaming ASR that corrects pronunciation in real time, TTS that reads the hard sentence back, and no account, no cloud, no data leaving the tablet. Parents notice that.
  • Voice control for an existing app. Dictation and read-aloud in AI Chat is the template — Google even open-sourced the reference implementation, including how they strip Markdown and skip code blocks before synthesis so the voice output doesn't read out backticks.

The cross-platform path is the one I'd watch longest. A 31M-parameter TTS model with sub-100 ms first audio on a web page is the kind of thing that makes voice interfaces cheap enough to put everywhere, not just in flagship apps.

The catches, stated plainly

  • The 0 MB claim is Android-only. It holds because the models live in the OS via AICore. On iOS, desktop, or web, you're shipping .litertlm bundles yourself, and those files have real sizes Google doesn't quote in the announcement.
  • Device support is not universal. AICore and NPU/TPU acceleration depend on the device. Older or lower-end Android phones won't get the same speed, and Google doesn't publish a minimum-spec list in this announcement.
  • No independent benchmarks yet. Every accuracy and latency figure above is Google's. The open-weight models (Parakeet, Kokoro) at least have public track records outside Google, but the ML Kit numbers are launch-day claims until someone reproduces them.
  • The models are small on purpose. A 100M-parameter TTS model and a 0.6B ASR model will not match a cloud model on hard cases — heavy accents, overlapping speakers, noisy rooms. The trade is latency and privacy for raw quality, and for some apps that's the wrong trade.
  • This is day one. The APIs are new, the docs are fresh, and the rough edges haven't been found by the community yet. If you ship on this tomorrow, you're the beta tester.

How to try it today

Download the latest Google AI Edge Gallery app on Android or iOS and try the XR Language Teacher and voice-enabled AI Chat — that's the fastest way to feel the latency for yourself. The reference implementation is open source on GitHub, and Google says you can point a coding agent at the repo to adapt the streaming ASR, bilingual TTS code-switching, and sentence-synchronized playback pipeline. MediaPipe's ASR and TTS tasks can be tried directly in the browser. And if you want to benchmark on real hardware, Google points to its Developer Device Platform with 200+ physical devices.

My honest take: the headline is "offline voice," but the real story is that voice just became a local feature instead of a cloud service. That changes the economics — no per-minute bill — and the privacy story — nothing leaves the phone — in ways that matter more than any single benchmark. The numbers will get checked by the community in the coming weeks. What won't change is that the cheapest, most private voice interface you can ship is now the one that never calls home.

Have you tried the Edge Gallery voice features yet? And if you have — does the offline transcription hold up in a noisy room, or is that where the small models show their limits?

Sources: Google Developers Blog, "On-Device Speech Recognition and Text-to-Speech in Google AI Edge Gallery," November 10, 2026. All capability, benchmark and performance claims are Google's, stated as such. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.