GoogleGoogle AI Edge Team, Android ML TeamAnnouncement
On-Device Speech Recognition and Text-to-Speech in Google AI Edge Gallery
Summary
Google AI Edge Gallery now includes on-device Automatic Speech Recognition (ASR) and Text-to-Speech (TTS), enabling offline voice interactions on phones and wearables. ML Kit for Android delivers multilingual ASR across 100+ languages at 10x real-time speed with 8-9% Word Error Rate on FLEURS, and a diffusion TTS that begins speaking in ~100 ms, synthesizes 24 kHz audio at up to 50x real-time speed, and cuts power by over 50%. For cross-platform apps, MediaPipe Tasks and LiteRT support open-weight models like Parakeet-TDT 0.6B, PocketTTS 100M, Kokoro-82M, and Phonon 31M, packaged as single .litertlm files. These models run alongside on-device small language models like Gemini Nano and Gemma 4, powering the XR Language Teacher showcase with bilingual TTS code-switching and voice-enabled AI Chat with sentence-synchronized playback. All building blocks, including ML Kit APIs, MediaPipe Tasks, and LiteRT model bundles, are available today for developers to integrate into their own apps.
Context
Voice interaction on mobile and wearable devices often requires a network connection for cloud-based speech processing. Google AI Edge Gallery now provides on-device ASR and TTS to enable fully offline, low-latency voice interactions using compact speech models that run alongside on-device small language models.
Approach / What changed
The release integrates system speech models via ML Kit for Android, offering multilingual ASR across 100+ languages and a diffusion TTS model. For cross-platform apps, MediaPipe Tasks and LiteRT support open-weight models like Parakeet-TDT 0.6B, PocketTTS 100M, Kokoro-82M, and Phonon 31M, packaged as .litertlm files. The pipeline powers the XR Language Teacher showcase with bilingual TTS code-switching and voice-enabled AI Chat with sentence-synchronized playback.
Takeaways
ML Kit’s on-device ASR achieves 8.0-9.0% Word Error Rate on FLEURS across 100+ languages and transcribes speech 10x faster than real time, with end-to-end speech translation wait times reduced by 35-44%.
The on-device diffusion TTS model begins speaking in ~100 ms, synthesizes 24 kHz audio at up to 50x real-time speed, and reduces active device power by over 50% compared to traditional autoregressive TTS.
For cross-platform apps, MediaPipe Tasks and LiteRT support open-weight models like Parakeet-TDT 0.6B, PocketTTS 100M, Kokoro-82M, and Phonon 31M, all packaged as single .litertlm files, with a growing catalog from the Hugging Face LiteRT Community.
FAQ
What languages does the on-device ASR support?
The on-device ASR supports multilingual transcription across 100+ languages, achieving 8.0-9.0% Word Error Rate on the FLEURS benchmark.
How fast is the on-device TTS?
The on-device TTS begins speaking in approximately 100 ms and synthesizes 24 kHz audio at up to 50x real-time speed, generating a 27-second paragraph in about 500 ms.
Can I use custom speech models with this pipeline?
Yes, LiteRT provides a rich set of multimodal model examples and supports custom implementations. The Hugging Face LiteRT Community offers a growing catalog of open-weight models like Whisper, Qwen3-ASR, and Moonshine for ASR, and Qwen3-TTS and Matcha for TTS.
Source and freshness
- Source
- developers.googleblog.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.