Skip to content

Google AI Edge Gallery Adds On-Device Speech and TTS

Google's AI Edge team shipped offline speech recognition and text-to-speech for Android via ML Kit and for cross-platform apps via MediaPipe Tasks and LiteRT.

Draftship Editorial4 min read

What Google announced

On Nov. 10, 2026, the Google AI Edge team said Google AI Edge Gallery now supports on-device speech recognition and TTS that runs fully offline. The announcement describes a local voice pipeline that pairs compact speech models with on-device small language models such as Gemini Nano and Gemma 4, so an app can stream live transcription while a user speaks and synthesize spoken replies on the device.

The release covers two build paths: ML Kit for Android, and MediaPipe Tasks plus LiteRT for cross-platform apps using open-weight models. Google also built a reference app, the XR Language Teacher, and added voice input and read-aloud playback to AI Chat inside the Gallery app.

How the Android path works

ML Kit's GenAI Speech Recognition and Text-to-Speech APIs rely on models managed through AICore and Google's on-device speech services. Google says this adds 0 MB to an app's APK, gives hardware acceleration across NPU and TPU backends where available, and updates models in the background.

Numbers from the source: transcription covers 100+ languages and reaches 8.0% to 9.0% word error rate on FLEURS, running about 10x faster than real time (roughly 300 ms per 3 seconds of speech) and cutting end-to-end speech translation waits by 35% to 44%. The TTS model is a few-step diffusion model accelerated on NPU that starts speaking in about 100 ms, produces 24 kHz audio across 8 voices and 21 locales at up to 50x real time, synthesizes a 27-second paragraph in about 500 ms, and uses over 50% less active device power than autoregressive TTS. It also supports earcons.

The cross-platform path

MediaPipe Tasks wraps the plumbing: audio framing, configurable voice activity detection and decoding. Its ASR task supports streaming, full-sentence batch and hybrid modes, and ships with Parakeet-TDT 0.6B as an example model; it runs on iOS, macOS, Windows, Linux and Web. The TTS task adds voice selection and low-latency streaming output, with PocketTTS 100M, Kokoro-82M and Phonon 31M (sub-100 ms time to first audio on most devices).

Beneath both, LiteRT packs a model's parts, including encoder and decoder subgraphs, tokenizers, phonemizer dictionaries, voice embeddings and vocoders, into a single .litertlm file. Ready-made bundles include Parakeet-TDT-0.6B-v3, PocketTTS 100M, Phonon 31M and Kokoro-82M, and Google points to the Hugging Face LiteRT Community for others such as Whisper, Qwen3-ASR, Moonshine, Qwen3-TTS and Matcha.

What it changes for people building apps

Voice features no longer have to mean a server round trip per utterance. The choice is mostly reach versus control. ML Kit gives Android a managed, OS-level route with no model downloads to handle, but it is Android only. MediaPipe Tasks and LiteRT reach more platforms and let you swap in open-weight models, at the cost of choosing, bundling and testing those models yourself. Both designs are meant to run alongside an on-device language model without competing for compute.

The example Google built

In XR Language Teacher, a user points a camera at an object or menu and asks a question aloud. ASR streams the question alongside the camera frame, in a single pass, to Gemini Nano on Android or Gemma 4 elsewhere. A structured output schema lets the model answer in the user's own language while tagging foreign phrases, so TTS can switch voices and accents mid-sentence and show a vocabulary card. Google says the same experience extends to smart glasses, with inference running on a paired phone.

Limits and open questions

The source gives no pricing, licensing terms or minimum Android version for the ML Kit GenAI APIs, and does not list which devices get NPU or TPU acceleration. It names Gemini Nano and Gemma 4 as companions but gives no model sizes or availability details for either. The performance figures are Google's own and are not backed by an independent benchmark here. Availability is stated as of the announcement, on Android and iOS for the Gallery app, though the ML Kit route described is Android specific.

What it means for AI app builders

Cheap, offline voice removes one common reason to send audio to a hosted API, which matters for privacy and for apps used where connectivity is poor. If you are wiring model calls into an app, Draftship's guide to adding AI to your app covers the application side. Draftship builds and hosts apps with a built-in database and user sign-in, and is not affiliated with Google.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.