The four-color microphone button sitting in the Google bar on most Android homescreens has done one thing for years: listen to your voice and hand you a standard web search. Starting with this week’s Google app beta, it does four things \u2014 and it remembers which one you used last.<\/p>\n
Google is testing a sweeping redesign of the voice search interface on Android that folds three previously separate AI-powered features into a unified entry point. The update, spotted in Google app versions 17.1 and 17.2 by 9to5Google’s Abner Li<\/a>, adds a toolbar at the bottom of the screen that lets users move between Search, AI Mode, Search Live, and Song Search without leaving the voice interface. It is not yet broadly available, but it signals how Google intends the homescreen voice button to work as it consolidates its AI search products.<\/p>\n Voice Search Gains a Mode Switcher \u2014 Borrowed From Lens<\/p>\n The redesign takes after the mode bar Google has already established in Google Lens, where users choose among Search, Translate, Live, and Create rather than hoping the app guesses their intent. The voice interface now adopts the same logic: instead of a single fixed function, the toolbar makes four modes explicitly chooseable.<\/p>\n Standard Search remains the baseline \u2014 speak a query, see real-time transcription at the top of the screen via a new four-bar waveform animation, and land on a traditional results page. The compact waveform animation that appeared in the earlier January 2026 visual refresh has been replaced by this four-bar design, which is consistent with the animated arcs already used in AI Mode and Search Live.<\/p>\n For AI Mode queries, users tap a next arrow to submit; a stop button on the left returns them to a search box containing a full transcript for further editing before searching. Search Live offers the same back-and-forth conversational flow. Song Search remains visually unchanged in function, displaying the “Play, Sing, Hum” invitation that replaced an older globe animation earlier this year.<\/p>\n The single most useful element in the redesign may be the smallest: the toolbar remembers whether a user last selected<\/a> Search or AI Mode and defaults to that choice the next time the voice interface opens. For a user who reaches for AI Mode every morning, the homescreen button will now open AI Mode without a manual tap to switch.<\/p>\n What Each Mode Actually Is \u2014 and the Different AI Powering Each<\/p>\n The four modes are not interchangeable variants of the same system. Each draws on a distinct AI architecture, built for a different user need.<\/p>\n AI Mode runs on Gemini 3.5 Flash, which Google made AI Mode’s default model<\/a> globally at I\/O 2026, and performs retrieval-augmented generation: it fetches content from multiple web sources in parallel, then synthesizes a new prose answer with inline citations rather than returning a list of links. This makes it well suited to complex multi-part questions where synthesis adds more value than a ranked list. Google launched AI Mode as a Search Labs experiment in March 2025, opened it to all US users in May 2025, and by Google I\/O in May 2026, AI Mode had crossed one billion monthly users<\/a> and was live in nearly 200 countries and 98 languages.<\/p>\n Search Live is powered by Gemini 3.1 Flash Live<\/a>, released on March 26, 2026, which Google describes as its highest-quality audio model to date. The model works differently from traditional voice AI in a technically significant way: it collapses the conventional three-step pipeline \u2014 transcribing speech to text, running that text through a language model, then synthesizing speech back \u2014 into a single native audio-to-audio process. The result is roughly 960-millisecond response latency with no intermediate transcription step, which is why Search Live conversations feel more fluid than earlier voice AI. The model uses WebSocket bidirectional streaming, meaning users can interrupt mid-response and the system adjusts. Google expanded Search Live globally<\/a> in March 2026, and it now operates in 90-plus languages across more than 200 countries. Users can also point their camera at an object during a Search Live session to ask questions about what the camera sees.<\/p>\n Song Search is technically distinct from both. It does not use Shazam-style acoustic fingerprinting, which matches a recording playing in the room against a database of stored audio hashes. Instead, Google’s system uses machine learning melody embedding<\/a> to strip an input \u2014 whether hummed, whistled, or sung \u2014 down to its melodic sequence, converts that sequence to a numerical embedding, and compares it against a database of more than 500,000 songs. The practical difference: Shazam needs the song playing out loud; Song Search works when the song exists only in the user’s memory. The model was trained on pairs of hummed and studio-recorded audio using triplet loss, teaching it to recognize a melody regardless of key, tempo, or vocal timbre.<\/p>\n Standard Search is unchanged: transcription into a traditional Google results page.<\/p>\n How Does Google AI Mode Compare to Standard Voice Search on Android?<\/p>\n AI Mode replaces the returned search page with a Gemini-synthesized prose answer drawing on multiple live web sources. Standard Search returns the familiar ranked link list. The key functional difference is that AI Mode’s answers are composed by the model \u2014 not just surfaced from the web \u2014 which makes them faster to read but one step further from the primary sources. The new toolbar makes this distinction explicit and user-chosen for the first time.<\/p>\n Why the Toolbar Design Matters \u2014 and What It Reveals About Mode Memory<\/p>\n The move to an explicit mode switcher, rather than a single-function button, is a design acknowledgment that Google’s voice entry point now routes to fundamentally different AI systems. The earlier single-function voice button was workable when voice meant “transcription to search.” It becomes harder to navigate when voice means “which of four distinct AI systems should interpret this?”<\/p>\n The mode-memory feature is notable beyond its convenience value. Each time a user selects AI Mode rather than standard Search from the homescreen, Google captures a behavioral signal that prior Android voice interactions did not generate: which interface mode the user prefers at the homescreen level. That preference data is generated at an entry point closer to the operating system than previous search behavior data, and it is now subject to Google’s June 2026 Search Services History setting<\/a>, which by default captures voice and media queries submitted through Search tools for AI model training. Users who prefer AI Mode consistently may find their homescreen behavior informing model training unless they specifically review that setting.<\/p>\n Where the Redesign Fits in Google’s Broader Platform Transition<\/p>\n The timing of this redesign is not arbitrary. Google Assistant was retired on March 31, 2026, with Gemini taking over as the unified AI layer across Android, Wear OS, Chrome, and smart speakers. The voice button on the Android homescreen is, for a large share of Android users, the most accessible voice entry point on their phone \u2014 and it is now being repositioned as the gateway to Google’s full AI search portfolio, not just a dictation shortcut.<\/p>\n At Google I\/O in May 2026, the company confirmed that Gemini 3.5 Flash now powers every Google Search query globally as the default, replacing the traditional ranked-link model with AI-generated answers. The homescreen voice toolbar extends that shift to the mobile widget level, placing AI Mode one tap away from any Android homescreen that carries the Google bar.<\/p>\n