Skip to content
AI.info

The Pulse

Google adds offline speech input and playback to AI Edge Gallery

Google AI Edge Gallery has added on-device speech recognition and text-to-speech for its AI Chat and language-teaching features. The company also detailed Android and cross-platform tools for developers building local voice applications.

Google adds offline speech input and playback to AI Edge Gallery

Google has added offline speech recognition and text-to-speech to AI Edge Gallery, giving users a way to dictate questions and hear responses from on-device models. The November 10 announcement on the Google Developers Blog also introduced a language-teaching showcase that combines a phone camera with spoken questions and answers.

The features are available in the latest AI Edge Gallery app for Android and iOS. The app pairs speech models with small language models such as Gemini Nano and Gemma 4, with processing taking place locally rather than requiring a network connection.

Voice controls in chat

In AI Chat, users can dictate a message and tap a speaker button on a response to hear it read aloud, or enable automatic playback. A floating playback bar highlights the sentence being spoken and lets users jump to another sentence by tapping it.

Model Management provides controls for choosing available speech recognition and text-to-speech pipelines, preset voices and downloaded language packs. The app's open-source reference code also demonstrates how to remove Markdown syntax from spoken responses and skip material such as code snippets and tables. It shows how to schedule speech synthesis by sentence to avoid competition for computing resources and memory while the language model generates tokens.

A camera-based language teacher

A separate showcase called XR Language Teacher combines the camera, language model and speech pipeline. A user can point a camera at an object, menu or sign and ask a question aloud. Speech recognition streams the question alongside the camera image to Gemini Nano on Android devices or Gemma 4 on other platforms.

The output schema tags foreign-language phrases separately, allowing the speech pipeline to switch voices and accents while displaying a vocabulary card. The showcase can also be adapted for hands-free smart glasses, with inference running locally on a paired phone.

Android speech models

For Android developers, Google offers speech recognition and text-to-speech APIs through ML Kit. The recognition API accepts audio from a device microphone or external streams, including smart glasses, and returns partial and final transcriptions as someone speaks.

Google says its recognition models support more than 100 languages and achieved an 8.0% to 9.0% word error rate on FLEURS. The company reports transcription at 10 times the speed of the audio being transcribed, or about 300 milliseconds for three seconds of speech.

For speech generation, the company says its model starts speaking in about 100 milliseconds and produces 24 kHz audio in eight voices across 21 locales, at up to 50 times real-time speed. Google says the models are integrated into the operating system and add 0 MB to an application's APK.

Tools for other platforms

For cross-platform applications, MediaPipe Tasks provides speech recognition and synthesis APIs across iOS, macOS, Windows, Linux and the web. Its speech recognition task supports streaming, full-sentence batch and hybrid transcription; the hybrid mode streams words as they are spoken and refines the sentence after speech ends.

The text-to-speech task includes voice selection and streaming audio output. The app ships with models packaged for LiteRT, including Parakeet-TDT-0.6B-v3 for multilingual recognition and PocketTTS 100M, Kokoro-82M and Phonon 31M for speech synthesis. Each model's components are packaged in a single .litertlm file, and developers who need more control over processing or custom model architectures can build directly on LiteRT.

Source

Companies in this story

Related stories

Explore

More articles