- Some DeepSeek app users have received a voice-interface test.
- The reported settings show four named voice options.
- The reports do not disclose the underlying speech architecture or a full-release date.
A visible app change for selected users
DeepSeek has begun a limited test of spoken interaction in its mobile application. IT Home reported on September 12 that participating users could see a speaker-shaped button in the upper-right corner of the app. TechNode subsequently described the ability to have responses read aloud.
The reported settings include four voice choices: Beike, Bailang, Haixing and Anchao. Their Chinese names refer to a shell, a white wave, a starfish and an undertow. The interface describes different vocal styles, ranging from lively to lower-pitched. These are selectable output voices, rather than four separately announced language models.
What is known about the rollout
The reports concern a staged test, often described in Chinese product coverage as a gray release. Access is limited to some users, and the sources do not give a date when every account will receive the same interface. An app update alone therefore does not establish that a particular user has been included.
The original reporting shows the controls and voice choices but does not explain how the speech system is built. It does not establish whether the implementation uses separate recognition and synthesis components, a single speech model, local inference or cloud processing.
An interface update rather than a model specification
The practical change is the availability of audible responses and a voice-oriented interaction surface. It belongs to the user-facing application layer. It should not be confused with an announcement of a public speech API or downloadable audio-model weights.
For now, the verified story is a limited product test with identifiable controls and voice settings. Claims circulating elsewhere about offline operation, zero latency or a fully local speech architecture go beyond the evidence in the original report and are not included here.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01Updates & corrections
— Expanded with additional reporting and source context.
