Google AI Edge Gallery adds on-device ASR and TTS, with ML Kit transcription across 100+ languages

Google has added on-device Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) to Google AI Edge Gallery, according to a Google Developers Blog post from the Google AI Edge and Android ML teams dated Nov. 10, 2026. The speech models run offline and are designed to pair with on-device small language models such as Gemini Nano and Gemma 4.

ML Kit on Android

Google says ML Kit now uses its latest on-device speech models. Both the ASR and TTS models are integrated into the OS, adding 0 MB to an app's APK. They are managed through AICore and Google's on-device speech services, so developers do not handle model downloads. Google says apps benefit from NPU and TPU acceleration when available, plus background model updates.

ML Kit GenAI Speech Recognition API

• Multilingual recognition across 100+ languages, reaching 8.0% to 9.0% Word Error Rate on FLEURS, according to Google.
• Transcribes at 10x real-time speed (about 300 ms per 3 seconds of speech).
• Cuts end-to-end speech translation wait times by 35% to 44%, Google says.
• Streams partial and final transcriptions from the built-in microphone or external audio streams such as smart glasses.

ML Kit GenAI Text-to-Speech API

• Uses a compact, few-step diffusion model accelerated on NPU.
• Begins speaking in about 100 ms and generates 24 kHz audio at up to 50x real-time speed.
• Offers 8 expressive voices across 21 locales, with built-in earcon support.
• Synthesizes a 27-second paragraph in about 500 ms.
• Cuts active device power by over 50% compared to traditional autoregressive TTS models, according to Google.

Cross-platform: MediaPipe Tasks and LiteRT

For apps outside the Android system path, Google optimized new .litertlm speech bundles: Parakeet-TDT 0.6B, PocketTTS 100M, Kokoro-82M and Phonon 31M. Google says they deliver real-time streaming transcription and low-latency speech synthesis in a compact memory footprint, and run alongside on-device SLMs without compute contention.

MediaPipe Tasks provides ASR and TTS APIs across iOS, macOS, Windows, Linux and Web. The ASR Task handles audio framing, configurable Voice Activity Detection (VAD) and decoding. It supports real-time streaming, full-sentence batch and hybrid modes, where words stream live and the full sentence is refined once speech ends.

Showcases in Google AI Edge Gallery

XR Language Teacher. The user points a camera at an object or menu and asks a question aloud. On-device ASR streams the question with the camera frame to Gemini Nano on Android, or Gemma 4 on other platforms, in a single pass. A structured output schema tags foreign-language phrases so the TTS pipeline can switch voices and accents mid-sentence, alongside an on-screen vocabulary card. Google says the same experience extends to smart glasses with inference running on a paired phone.

AI Chat. ASR and TTS are now built into the main AI Chat feature. Features include:

• ASR and TTS settings in Model Management, covering pipelines, preset voices and language packs.
• An "Auto-play TTS" option and a speaker button that opens a floating bar highlighting each spoken sentence, with tap-to-jump.
• An open-source reference implementation that strips Markdown, skips code snippets and tables, and schedules synthesis per sentence to avoid compute and memory contention during LLM token generation.

Google notes that multimodal models like Gemma 4 can process audio directly, but it offers a modular cascade of separate lightweight voice models for latency optimization and flexibility. Open-source code on GitHub shows the recommended path for each platform.

What this means

Google is offering two routes. On Android, ML Kit's system-integrated models add 0 MB to an APK and are managed through AICore. Elsewhere, MediaPipe Tasks and LiteRT supply open-weight speech bundles from 31M to 0.6B parameters. The reference implementation's sentence-level scheduling addresses a practical constraint Google names directly: speech synthesis competing with LLM token generation for compute and memory. Google's stated reason for separate voice models alongside multimodal models like Gemma 4 is latency optimization and flexibility.