Starting September 12, users began noticing a small speaker icon in the top-right corner of the DeepSeek app. Tapping it opens a voice chat interface. Accounts that got into the test can also see a new "reading voice" option under the voice settings tab.
There are four voices on offer, each with a short name and description: Beike (Shell, playful and lively), Bailang (White Wave, bright and steady), Haixing (Starfish, cute and sweet), and Anchao (Undercurrent, deep and resonant). All four names are drawn from the sea.
There has been no official announcement for this rollout — the news has spread entirely through user reports. DeepSeek has not disclosed the grayscale ratio, which devices are covered, or when the feature might reach general availability.
Two events, two days apart
Two days earlier, on September 10, DeepSeek released V4.1 Flash, the smallest model in its new architecture series, with native multimodal vision understanding. That release followed the standard playbook: an announcement, technical documentation, the works.
The two moves came within days of each other but couldn't be more different in approach. The model side was public, documented, and verifiable; the app side was pushed quietly, with no statement, leaving users to discover it on their own. This split is consistent with how DeepSeek typically operates: models are things to announce, products are things to ship first and talk about later — if at all.
Functionally, what's rolling out is a read-aloud-plus-chat combination, still some distance from the real-time voice dialogue that's becoming the norm elsewhere. The term "reading voice" itself points to text-to-speech — the model finishes writing a response, then reads it aloud — rather than the interrupt-capable, think-while-listening full-duplex interaction some competitors already offer. Whether a fuller voice mode is queued up behind this one isn't yet clear.
What this means for users in China
First, access. During grayscale testing, the only way to check is to open the app and look for the speaker icon in the top-right corner. If it isn't there, your account hasn't been selected yet — reinstalling the app or switching accounts won't help. This kind of staged rollout is common among domestic apps: it limits the blast radius if something goes wrong, at the cost of giving users no sense of when their turn will come.
Second, competitive positioning. Several major domestic AI apps shipped voice features well before DeepSeek — Doubao, Tongyi, Yuanbao, and Kimi have all had voice chat running for a while, and some already support real-time interruption. DeepSeek is catching up here, closing a gap in product completeness rather than breaking new ground.
Third, how the use case shifts. DeepSeek's user base skews toward long-form text, coding, and reasoning tasks — scenarios where voice doesn't add much value by default. Having code read aloud isn't particularly useful, and listening to a long chain of reasoning is arguably more tedious than reading it. Voice is more likely to find its footing in two narrower situations: listening to results while commuting or doing chores, and dictating a question when typing isn't convenient. The four voice options look more like a tuning knob for listening preference than a new interaction mode.
Fourth, what this update signals. DeepSeek's product cadence has historically lagged its peers, with the app's feature set long limited to a basic chat box. Voice entering grayscale testing, paired with V4.1 Flash's native vision understanding, suggests the app's capability set is expanding toward multimodal use. How far that expansion goes, and whether it eventually reaches real-time interaction, is something DeepSeek hasn't indicated either way.
For anyone hoping to use voice chat right now, there's really only one piece of advice: wait. There's no application process for the grayscale test, and no public queue to join.
Sources: IT Home, Sina Technology, CocoLoop, DeepSeek app settings page; verified the four voice names and descriptions, the grayscale start date, and the V4.1 Flash release timeline.