Google AI Edge Gallery voice: offline speech and device limits
Google AI Edge Gallery now has a documented voice workflow for talking to local models and hearing their replies. The useful distinction is not simply offline versus cloud: the Android system APIs, downloadable open-weight speech models and Gallery app have different compatibility and setup requirements.
Google described the integration in a November 10 developer announcement. The November 12 Hugging Face community article republishes that announcement. This is not a November 12 model launch: Gallery 1.0.23 release notes, dated November 4, already mention speech stability improvements and XR Language Teacher activities.
What the voice workflow does
The announcement describes streaming speech input and spoken responses in AI Chat, plus an XR Language Teacher demonstration that combines a camera image, a question and a bilingual answer. Sentence highlighting and playback controls let a user follow a spoken reply. These are Google’s descriptions, not ByteForward hands-on results.
The Gallery repository lists Android 12 and later and iOS 17 and later as app requirements, with Google Play and App Store download links. It calls Gallery an experimental beta. Meeting that app-level OS requirement does not establish support for every speech engine or Gemini Nano feature.
ML Kit support is narrower than the app’s OS requirement
The ML Kit speech-recognition documentation separates Basic mode from Advanced mode. Basic runs on Android API level 31 and above. Advanced depends on selected Gemini Nano devices, including Pixel 10 and Pixel 11 series and listed Snapdragon, Dimensity and Exynos chipsets. The SDK’s minimum API level of 26 is not the runtime compatibility list.
Recognition produces text in the spoken language. Translation requires the separate Speech Translation API. Developers should also distinguish partial transcriptions, which can change, from final results. The recognition API is alpha, with no service-level agreement or deprecation policy; Google warns that backward-incompatible changes are possible.
The separate ML Kit text-to-speech API is also alpha. Its supported-device list includes Pixel 8 through Pixel 11 and specified Samsung Galaxy models. It requires an up-to-date Google Speech Recognition & Synthesis app. Voice availability must be checked, and a voice may need downloading before synthesis works. Google documents eight voices and 21 locales, not universal language coverage.
Open-weight bundles make a different tradeoff
Developers targeting other platforms can inspect the downloadable speech bundles rather than assume Android’s system models are available everywhere. The PocketTTS conversion card lists a 156 MB English bundle and a 725 MB seven-language bundle. Those are file sizes, not a promise about total runtime memory. The packages combine speech-model components and voice assets for MediaPipe and LiteRT.
The Parakeet-TDT conversion card limits its language coverage to 25 European languages. Its documented streaming approach can expose provisional words before confirming boundary-spanning speech. That is a practical reason to avoid treating every live word as a finished transcript.
Gallery’s code license is Apache 2.0, but that does not replace model licenses. The inspected PocketTTS package is labeled CC BY 4.0; the Parakeet card likewise identifies CC BY 4.0 for weights and tokenizer and separate Apache-licensed conversion and runtime components.
Offline inference still needs preparation
The local speech path can avoid sending inference to a remote model, but downloads, service initialization and updates still require connectivity. ML Kit’s documentation explicitly includes downloadable and unavailable states. Prepare the required models and voices before relying on the app without a connection.
For a deployment decision, check the actual device, selected engine, language and model license together. Local execution is not evidence that every optional Gallery skill is offline, nor a substitute for checking an application’s permissions and data handling.
For a different local-model approach, our LFM3 deployment guide separates audio-capable checkpoints from text models and explains why runtime support and model licensing need their own checks.
Image: AI-generated editorial illustration, not a product interface. Google and Google AI Edge are trademarks of Google LLC. ByteForward is not affiliated with or endorsed by Google.
