On-device open-weight speech and deployment

Google AI Edge Gallery brings open-weight speech models into offline voice apps

Google outlines local ASR and TTS stacks around Parakeet, PocketTTS, Phonon and Kokoro, with separate Android-system and cross-platform model paths.

Open Model Weights published15 Nov 2026
Primary sourceGoogle AI Edge Team / Hugging Face — On-Device Speech Recognition and Text-to-Speech in Google AI Edge Gallery
Source published2026-11-12
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original high-gloss CGI-style editorial art of a glass smartphone with luminous turquoise sound waves, polished titanium frame and floating local speech-processing rings.
Original cinematic conceptual rendering of offline speech processing on a personal device; not an actual Google AI Edge Gallery screenshot or device performance measurement. AI-generated editorial illustration; not a documentary photograph.

Google's AI Edge team has described a broader offline voice stack for AI Edge Gallery, combining local speech recognition and text-to-speech with small on-device language models. The November 12 technical post describes both Android-integrated speech services and cross-platform pathways built around MediaPipe Tasks and LiteRT. For open-weight deployment, the distinction matters: open speech checkpoints can be packaged as portable model bundles, while some Android speech components remain managed through device system services rather than offered as interchangeable public weights. Primary source ↗

The cross-platform examples are especially practical. Google names Parakeet-TDT-0.6B-v3 for multilingual automatic speech recognition, PocketTTS 100M and Kokoro-82M for multilingual speech synthesis, and Phonon 31M for a lightweight English TTS path. The downloadable LiteRT model bundles package the components needed for deployment, reducing application integration complexity. MediaPipe exposes speech tasks that handle audio framing, voice activity detection and streaming, while LiteRT provides lower-level inference support.

The accompanying demos connect local speech to an on-device assistant: live spoken questions can be transcribed, passed into a small multimodal language model and answered aloud without sending ordinary inference traffic to a cloud service. The article describes an XR Language Teacher showcase and voice input/output in AI Chat. Google reports low time-to-first-audio and strong language coverage for its optimised pipelines, but those figures depend on devices, system components and specific model selections. We treat them as publisher-reported, not independently verified.

For the Open Model Weights registry, this is an integration milestone rather than a single new model announcement. Model identity, weight access, license and benchmark claims should be checked individually for Parakeet, Phonon, PocketTTS and Kokoro. Keeping the boundary between public downloadable model bundles and Android-managed system models explicit helps readers choose a genuinely portable offline solution without assuming that every part of the showcase is itself open weight.

OMW REGISTRY WATCH

What this changes for the evidence layer

This is a model-integration and packaging release, not a single newly trained open-weight foundation checkpoint. Differentiate publicly downloadable LiteRT/MediaPipe speech bundles from Google's managed Android-system models; check artifact license, revision, device support and acoustic performance per bundled model before recording a model-specific capability.

PRIMARY SOURCE

Google AI Edge Team / Hugging Face — On-Device Speech Recognition and Text-to-Speech in Google AI Edge Gallery

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗