Google introduced on-device automatic speech recognition (ASR) and text-to-speech (TTS) in AI Edge Gallery on November 10, 2026, alongside reference code for local voice applications. The release combines an Android system-model route through ML Kit with a separate open-weight route through MediaPipe Tasks and LiteRT, according to the Google announcement.
The November 12 Hugging Face article explicitly identifies itself as a republication of the November 10 original. Google described the building blocks as available, rather than announcing a future release date.
What changed in AI Edge Gallery
The Gallery repository now describes microphone dictation and spoken responses in AI Chat, with selectable speech engines, voices and languages. Its XR Language Teacher combines camera, voice and text input with spoken answers and vocabulary cards. The repository still labels Gallery an experimental beta.
The published reference implementation also demonstrates sentence-level speech playback, including skipping code and tables when reading model responses aloud. This adds a concrete implementation reference for university developers evaluating voice interfaces alongside existing speech-recognition workflows.
Two deployment routes, with different constraints
The implementation choices carry different dependencies. The table combines Google’s release description with its API documentation; a capability in one route should not be assumed to apply to the other.
| Route | What the sources establish | Deployment qualification |
|---|---|---|
| ML Kit on Android | System speech models managed through AICore and on-device speech services. | The speech-recognition API is alpha; Advanced mode supports selected devices. |
| MediaPipe Tasks | Audio Transcriber supports complete clips and streaming audio, with timestamped transcript segments. | Interim transcripts can change before final results arrive. |
| LiteRT model bundles | Open-weight speech models package their components in .litertlm files. | This is a model-bundle route, distinct from Android-managed system models. |
The ML Kit documentation warns that its alpha API has no service-level agreement or deprecation-policy protection and may introduce incompatible changes. It also instructs developers to check model availability and download a model if needed. Its recognition API returns text in the spoken language; translation uses a separate API.
What the release means for research teams
The MediaPipe documentation says Audio Transcriber processes audio locally without sending audio or transcripts off the device. Its browser demo caches model weights after an initial download. Offline inference therefore needs to be distinguished from initial provisioning.
For researchers evaluating interview transcription, the distinction between interim and final text matters when preserving research records. A local voice prototype still needs an explicit choice of transcription conventions and a way to check the resulting transcript against the recording.
Google reports an 8–9% word error rate on FLEURS for its system ASR and about 100 milliseconds before its system TTS starts speaking. These are company-reported measurements, not evidence of accuracy on a particular university’s interview corpus or identical performance across devices.
For research offices, the practical opportunity is to evaluate local processing and inspectable reference code. Our assessment is that processing location alone does not establish participant consent, storage arrangements or access controls for the surrounding application. The announcement supplies implementation options, not an institutional approval for handling research recordings.
Sources
- Google AI Edge Team and Android ML Team: On-Device Speech Recognition and Text-to-Speech in Google AI Edge Gallery, November 10, 2026. Checked 2026-11-15.
- Hugging Face LiteRT Community: republication of the Google announcement, November 12, 2026. Checked 2026-11-15.
- Google AI Edge Gallery repository and README. Checked 2026-11-15.
- Google ML Kit: GenAI Speech Recognition API. Checked 2026-11-15.
- Google MediaPipe: Audio Transcriber. Checked 2026-11-15.
