SI Agent News
DevelopingGeminiOpen modelsResearch

Google adds offline speech tools to AI Edge Gallery

Google released on-device speech recognition and text-to-speech tools, providing developers with two distinct paths for building local voice applications.

Blueprint drawing: picture for the story 'Google adds offline speech tools to AI Edge Gallery'
AI-made blueprint drawing for this story. No real people or products are shown.
8–9%word error rate for system ASR

The short version

Why it matters

This release allows developers to build voice-enabled apps that function without an internet connection, which is critical for privacy and low-latency needs. It provides a concrete reference for researchers to evaluate how local processing impacts data handling and user experience.

New local voice capabilities

Google officially introduced on-device automatic speech recognition (ASR) and text-to-speech (TTS) within its AI Edge Gallery on November 10, 2026. These tools allow applications to convert spoken language into text and vice versa without relying on cloud servers. The release includes reference code designed to help developers build local voice applications more quickly.

The update specifically integrates microphone dictation and spoken responses into features like AI Chat. It also includes a specific example called the XR Language Teacher, which combines camera, voice, and text inputs with spoken answers. This repository currently functions as an experimental beta, providing a sandbox for testing these new capabilities.

One of the primary goals of this release is to provide a concrete implementation reference for university developers. These researchers can now evaluate voice interfaces alongside existing speech-recognition workflows. By providing these building blocks, Google is moving toward a more accessible framework for local audio processing.

Two distinct deployment routes

Developers choosing to implement these tools face two different routes, each with its own set of dependencies and constraints. The first is an Android system-model route that goes through ML Kit. This path is governed by specific API documentation and may require users to download models depending on availability.

The second option is an open-weight route that utilizes MediaPipe Tasks and LiteRT. This route is often preferred by those who need more control over the underlying models. As noted in the documentation, a capability present in one route does not automatically apply to the other, requiring developers to be careful during selection.

The ML Kit route currently uses an alpha API, which means it lacks a service-level agreement or a formal deprecation policy. This could lead to incompatible changes in the future. Consequently, developers must carefully weigh the stability of the system-model route against the flexibility of the open-weight MediaPipe route.

Privacy and local processing

A significant technical detail for researchers is how the MediaPipe Audio Transcriber handles data. This tool processes audio locally, meaning it does not send audio files or transcripts off the device to a remote server. While the browser demo caches model weights after an initial download, the actual inference remains on the local hardware.

For research teams, particularly those conducting interviews, the distinction between interim and final text is important for preserving records. Because the processing is local, it offers a path toward more private data handling. However, local processing alone does not automatically grant institutional approval for how data is stored or accessed.

Researchers still need to make explicit choices regarding transcription conventions. They must also establish a way to check the resulting transcripts against the original recordings. The tools provide the infrastructure for local processing, but the responsibility for data governance remains with the application developer.

Performance and benchmarks

Google has provided specific performance metrics to help developers understand what to expect from these tools. The company reports that its system ASR achieves an 8–9% word error rate on the FLEURS dataset. These numbers serve as a baseline for the capabilities of the current system-model route.

Regarding text-to-speech, Google reports a delay of about 100 milliseconds before the system begins speaking. These measurements are company-reported figures rather than evidence of how the tools will perform on every specific device. Performance may vary significantly depending on the hardware being used for the local inference.

It is important for developers to recognize that these benchmarks are not guaranteed for every use case. For example, accuracy might differ when used on a specific university interview corpus compared to the standard FLEURS dataset. Developers should test the tools on their specific hardware to determine real-world reliability.

What comes next for developers

The release of these building blocks marks a shift toward more modular voice integration in mobile and edge computing. Developers can now begin exploring how to skip non-essential elements, such as code or tables, when reading model responses aloud. This level of control is essential for creating natural-sounding interactive experiences.

Moving forward, the focus will likely shift toward optimizing these models for a wider variety of devices. Since the Gallery is currently in beta, we can expect more refinements to the API stability and model availability. Developers will need to stay updated on the ML Kit alpha status to avoid breaking changes.

For research offices, the immediate opportunity lies in evaluating the inspectable reference code. By using these tools, teams can prototype local voice interfaces while maintaining a clear view of how the underlying models function. This transparency is a key step for integrating AI into sensitive environments.

What to watch

Try it

Check the Google AI Edge Gallery repository to explore the reference code for the XR Language Teacher.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by CASRAI · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news