# Google AI Edge Gallery adds on-device speech recognition and text-to-speech

published: 2026-11-10T18:05:33.213Z
topic: infrastructure
source: https://developers.googleblog.com/on-device-speech-recognition-and-text-to-speech-in-google-ai-edge-gallery/
url: https://aiagentnews.io/stories/google-ai-edge-gallery-adds-on-device-speech-recognition-and-text-to-speech

Google has added on-device automatic speech recognition and text-to-speech to Google AI Edge Gallery, which the company says can run offline. The release covers ML Kit speech APIs for Android and open-weight speech models available through MediaPipe Tasks and LiteRT for cross-platform apps.

Google AI Edge Gallery now includes on-device automatic speech recognition (ASR) and text-to-speech (TTS), which Google says can run fully offline within a small memory footprint. The speech models are designed to pair with small language models such as Gemini Nano and Gemma 4. A new XR Language Teacher showcase combines the camera, speech and language model, and AI Chat gains voice input and read-aloud playback with sentence highlighting.

For Android, ML Kit's GenAI Speech Recognition and Text-to-Speech APIs run models managed through AICore, so apps do not need to bundle them. Google reports a word error rate of 8.0% to 9.0% on FLEURS across more than 100 languages, and says its TTS begins speaking in about 100 ms, outputs 24 kHz audio and supports 8 voices in 21 locales.

For cross-platform apps, MediaPipe Tasks provides speech recognition and synthesis interfaces for iOS, macOS, Windows, Linux and Web, built on LiteRT. Models including Parakeet-TDT 0.6B, PocketTTS 100M, Kokoro-82M and Phonon 31M ship as .litertlm bundles, and the Gallery code is open source on GitHub.

## Why it matters if you build agents

Developers building voice agents on Android can add streaming transcription and spoken replies without managing model downloads or making network calls for speech. Cross-platform apps can use the same open-weight models through MediaPipe Tasks or LiteRT, but they take on audio framing, voice activity detection and model selection themselves. The performance figures are Google's own, so measuring latency on the target devices is worth doing before committing to a design.