You remember the scene in a video, but not its filename or anything anyone said. A transcript search has very little to work with.
Google's October 6 release of EmbeddingGemma 2 gives developers another route: match a description against the pictures and sound themselves. The model turns text, images, audio and video into comparable numerical representations, so a spoken request can be used to look for a video clip. The query and the recording don't have to arrive in the same format.
Google points to a concrete example in its AI Edge Gallery app: Video Moments Finder, which locates moments using text or audio queries. A separate media-search example accepts an image as the search query. You can start with something that looks like what you remember, rather than first finding the words for it.
The model is designed to run on phones and laptops. Google says developers can build this search entirely offline, keeping the media processing on the device. That makes a private collection of recordings a possible search library without first uploading it to a cloud service.
These are similarity matches, so the result still needs watching. But the thing you remember can now be the starting point for the search: a scene, a sound or another picture, rather than a filename.
Sources:
https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2
Image: AI-generated editorial illustration of a personal media collection, not a Google product photograph or app screenshot.
Google's October 6 release of EmbeddingGemma 2 gives developers another route: match a description against the pictures and sound themselves. The model turns text, images, audio and video into comparable numerical representations, so a spoken request can be used to look for a video clip. The query and the recording don't have to arrive in the same format.
Google points to a concrete example in its AI Edge Gallery app: Video Moments Finder, which locates moments using text or audio queries. A separate media-search example accepts an image as the search query. You can start with something that looks like what you remember, rather than first finding the words for it.
The model is designed to run on phones and laptops. Google says developers can build this search entirely offline, keeping the media processing on the device. That makes a private collection of recordings a possible search library without first uploading it to a cloud service.
These are similarity matches, so the result still needs watching. But the thing you remember can now be the starting point for the search: a scene, a sound or another picture, rather than a filename.
Sources:
https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2
Image: AI-generated editorial illustration of a personal media collection, not a Google product photograph or app screenshot.