Posted by Terence Zhang – Developer Relations Engineer and Lisie Lillianfeld – Product Supervisor
TalkBack is Android’s display screen reader within the Android Accessibility Suite that describes textual content and pictures for Android customers who’ve blindness or low imaginative and prescient. The TalkBack group is all the time working to make Android extra accessible. Immediately, because of Gemini Nano with multimodality, TalkBack robotically gives customers with blindness or low imaginative and prescient extra vivid and detailed picture descriptions to higher perceive the photographs on their display screen.
Rising accessibility utilizing Gemini Nano with multimodality
Advancing accessibility is a core a part of Google’s mission to construct for everybody. That’s why TalkBack has a characteristic to explain pictures when builders didn’t embody descriptive alt textual content. This characteristic was powered by a small ML mannequin known as Garcon. Nonetheless, Garcon produced brief, generic responses and couldn’t specify related particulars like landmarks or merchandise.
The event of Gemini Nano with multimodality was the right alternative to make use of the newest AI know-how to extend accessibility with TalkBack. Now, when TalkBack customers choose in on eligible units, the display screen reader makes use of Gemini Nano’s new multimodal capabilities to robotically present customers with clear, detailed picture descriptions in apps together with Google Images and Chrome, even when the machine is offline or has an unstable community connection.
“Gemini Nano helps fill in lacking data,” stated Lisie Lillianfeld, product supervisor at Google. “Whether or not it’s extra particulars about what’s in a photograph a pal despatched or the type and reduce of clothes when procuring on-line.”
Going past primary picture descriptions
Right here’s an instance that illustrates how Gemini Nano improves picture descriptions: When Garcon is introduced with a panorama of the Sydney, Australia shoreline at night time, it’d learn: “Full moon over the ocean.” Gemini Nano with multimodality can paint a richer image, with an outline like: “A panoramic view of Sydney Opera Home and the Sydney Harbour Bridge from the north shore of Sydney, New South Wales, Australia.”
“It is superb how Nano can acknowledge one thing particular. As an illustration, the mannequin will acknowledge not only a tower, however the Eiffel Tower,” stated Lisie. “This type of context takes benefit of the distinctive strengths of LLMs to ship a useful expertise for our customers.”
Utilizing an on-device mannequin like Gemini Nano was the one possible answer for TalkBack to offer robotically generated detailed picture descriptions for pictures, even whereas the machine is offline.
“The typical TalkBack person comes throughout 90 unlabeled pictures per day, and people pictures weren’t as accessible earlier than this new characteristic,” stated Lisie. The characteristic has gained optimistic person suggestions, with early testers writing that the brand new picture descriptions are a “sport changer” and that it’s “great” to have detailed picture descriptions constructed into TalkBack.
Balancing inference verbosity and pace
One vital determination the Android accessibility group made when implementing Gemini Nano with multimodality was between inference verbosity and pace, which is partially decided by picture decision. Gemini Nano with multimodality at present accepts pictures in both 512 pixels or 768 pixels.
“The 512-pixel decision emitted its first token nearly two seconds sooner than 768 pixels, however the output wasn’t as detailed,” stated Tyler Freeman, a senior software program engineer at Google. “For our customers, we determined an extended, richer description was definitely worth the elevated latency. We have been in a position to conceal the perceived latency a bit by streaming the tokens on to the text-to-speech system, so customers don’t have to attend for the total textual content to be generated earlier than listening to a response.”
A hybrid answer utilizing Gemini Nano and Gemini 1.5 Flash
TalkBack builders additionally applied a hybrid AI answer utilizing Gemini 1.5 Flash. With this server-based AI mannequin, TalkBack can present the very best of on-device and server-based generative AI options to make the display screen reader much more highly effective.
When customers need extra particulars after listening to an robotically generated picture description from Gemini Nano, TalkBack provides the person an choice to hearken to extra by working the picture by means of Gemini Flash. When customers deal with a picture, they will use a three-finger faucet to open the TalkBack menu and choose the “Describe Picture” choice to ship the picture to Gemini 1.5 Flash on the server and get much more particulars.
By combining the distinctive benefits of each Gemini Nano’s on-device processing with the total energy of cloud-based Gemini 1.5 Flash, TalkBack gives blind and low-vision Android customers a useful and informative expertise with pictures. The “describe picture” characteristic powered by Gemini 1.5 Flash launched to TalkBack customers on extra Android units, so much more customers can get detailed picture descriptions.
Compact mannequin, huge influence
The Android accessibility group recommends builders trying to make use of the Gemini Nano with multimodality prototype and take a look at on a strong, server-side mannequin first. There builders can perceive the UX sooner, iterate on immediate engineering, and get a greater thought of the very best high quality attainable utilizing essentially the most succesful mannequin accessible.
Whereas Gemini Nano with multimodality can embody lacking context to enhance picture descriptions, it’s nonetheless finest follow for builders to offer detailed alt textual content for all pictures on their apps or web sites. If the alt textual content shouldn’t be supplied, TalkBack can assist fill within the gaps.
The Android accessibility group’s purpose is to create inclusive and accessible options, and leveraging Gemini Nano with multimodality to offer vivid and detailed picture descriptions robotically is a giant step in the direction of that. Moreover, their hybrid strategy in the direction of AI, combining the strengths of each Gemini Nano on machine and Gemini 1.5 Flash within the server, showcases the transformative potential of AI in selling inclusivity and accessibility and highlights Google’s ongoing dedication to constructing for everybody.
Get began
Be taught extra about Gemini Nano for app improvement.
This weblog publish is a part of our sequence: Highlight Week on Android 15, the place we offer assets — weblog posts, movies, pattern code, and extra — all designed that will help you put together your apps and benefit from the newest options in Android 15. You’ll be able to learn extra within the overview of Highlight Week: Android 15, which can be up to date all through the week.