Google’s Nov. 10 developer post describes a local speech stack for apps that need to listen and speak without sending audio to a cloud model. It pairs automatic speech recognition (ASR) and text-to-speech (TTS) with smaller on-device language models, so an app can stream a transcript and generate a spoken response while offline. The post presents the capabilities through Google AI Edge Gallery and explains how developers can integrate the underlying APIs and models.
Which integration path fits Android?
For Android, Google points to ML Kit GenAI Speech Recognition and Text-to-Speech APIs. The company says these use models managed through Android’s AICore and its on-device speech services, avoiding a separate model download in the app package. Google’s post describes partial and final streaming transcripts, microphone or external audio input, and speech output with selectable voices. Availability and performance can depend on compatible devices and the system model services, so teams should check the API requirements before designing an app around them.
Google reports transcription across more than 100 languages at up to 10 times real time, and says its speech model scored 8–9% word error rate on FLEURS. It also reports TTS starting in about 100 milliseconds with eight voices across 21 locales. These numbers come from Google’s announcement and test setup, not independent VIBPS measurements; actual results will vary with device, language, microphone and environment.
What about iOS, desktop and web?
For cross-platform projects, Google describes MediaPipe Tasks as a higher-level route for ASR and TTS on iOS, macOS, Windows, Linux and web. The tasks handle work such as audio framing, voice activity detection and decoding, and can run open-weight speech models. LiteRT is the lower-level inference layer for teams that want finer control or need to package supported model assets themselves. Those options trade some of ML Kit’s managed Android setup for more choices over models and platform integration.
Google lists examples including Parakeet-TDT for recognition and PocketTTS, Kokoro and Phonon for synthesis. Its post says bundled .litertlm files can package model components such as tokenizers, vocoders and voice embeddings. A compact bundle may simplify deployment, but it does not remove the need to measure memory, startup time, battery use, language coverage and accuracy on target hardware.
Is the whole experience private and offline?
The showcased voice pipeline runs on-device and can function without a network connection, according to Google. Whether an app stays fully offline depends on its own architecture: a developer could still send transcripts, prompts or user records to a server elsewhere in the product. Local inference also means the app must request and handle microphone permissions, provide a clear recording state and test what happens when a model or language pack is unavailable.
The new stack is useful for voice-first apps that need low latency, offline access or local handling of audio. The choice is between a managed Android path and more configurable cross-platform model tooling, rather than one identical SDK for every device. VIBPS has not benchmarked these APIs. Makers should test recognition quality and speech output with real speakers, noise, accents and interruption cases before promising a hands-free experience.
How Cursor’s agent voice chat differs from dictation
VIBPS is the social network for software. Makers can publish apps built with on-device AI and let people follow their work.