AI
Google Brings Speech Recognition and Voice Synthesis to On-Device AI Models
Google has released on-device automatic speech recognition and text-to-speech capabilities in its AI Edge Gallery, enabling voice interactions with mobile and wearable apps that work entirely offline.
Google Developers Blog (AI)
Google has unveiled new on-device speech capabilities that make it possible for mobile and web applications to perform automatic speech recognition (ASR) and text-to-speech (TTS) synthesis without relying on cloud services. These lightweight models, now available through Google AI Edge Gallery, enable developers to build voice-first applications that work offline with minimal latency and memory overhead.
The shift toward on-device speech processing addresses a key limitation of earlier AI systems: they required internet connectivity and cloud infrastructure to handle voice interactions. Modern compact speech architectures now make it feasible to run sophisticated multilingual ASR and expressive TTS models directly on phones and wearables, preserving user privacy and enabling seamless operation in areas with poor connectivity.
Platform-Specific Implementation Paths#
Android users benefit from tightly integrated speech models built into ML Kit. Google's latest on-device models deliver accurate transcription across more than 100 languages at 10 times real-time speed—meaning a 3-second audio clip transcribes in roughly 300 milliseconds. The accompanying TTS model, based on a few-step diffusion approach, begins producing audio in approximately 100 milliseconds and can synthesize 24 kHz speech at speeds up to 50 times faster than playback. Crucially, these capabilities add zero bytes to app install sizes, as the models are integrated into the operating system itself.
For cross-platform and web applications, Google has optimized a suite of open-weight model bundles compatible with MediaPipe Tasks and LiteRT, including Parakeet-TDT 0.6B, PocketTTS 100M, Kokoro-82M, and Phonon 31M. These models deliver real-time streaming transcription and low-latency synthesis while remaining lightweight enough to run alongside on-device small language models like Gemini Nano or Gemma 4 without resource conflicts.
Real-World Application: Language Learning#
To demonstrate the potential of locally-run speech processing, Google developed an XR Language Teacher showcase within the Edge Gallery. The application lets users point a phone's camera at unfamiliar objects, street signs, or menus and ask questions aloud—entirely offline. The system streams the user's spoken question in real time to Gemini Nano (on Android) or Gemma 4 (on other platforms), then generates a bilingual response. The TTS engine switches between the user's native language and the target language mid-sentence, with synchronized on-screen vocabulary cards showing relevant translations.

