GitHub Repo
Open WebUI Marks Residual Noise in Realtime Voice as Complete; Fix in a Release Still Unconfirmed
On October 14, Open WebUI marked a report about residual noise in Realtime voice as completed. The report says blank transcriptions may still leave audio audible to the model. The public issue lists no fix commit or release, so deployers should verify whether the server-side conversation is cleared as well.
Open WebUI’s Realtime voice feature has a report about conversation state synchronization: background noise may remain in the voice model’s context even when it produces no text message. User Classic298 filed the issue on October 12 and marked it completed on October 14. The public page does not provide a corresponding fix commit, so the confirmed change is the issue’s status; that alone does not establish that a released version has been fixed. Issue #32250
The reporter says that in v0.12.0 and the development build at the time, noise such as blowing into a microphone could be classified as speech while producing an empty transcription. The interface saved no message and gave no answer, but the audio segment remained in the provider-side conversation and continued to affect later turns. To reproduce the issue, generate noise several times during a Realtime call, then ask a normal question. The confused responses were the reporter’s observation; the page provides no quantitative evaluation or independent reproduction results. Reproduction steps
This case concerns the relationship between two conversation records. According to the official documentation, in Realtime mode the voice model listens and responds directly. The actual task is passed via a tool call to the model selected in the chat interface; its answer is written to the chat history and then read aloud by the voice model. In other words, visible text and the content received by the voice model belong to separate processing steps. Given this architecture, checking only whether an empty message appears in the chat window is not enough to verify that the call context is clean. Official voice mode documentation
For engineers maintaining voice agents, this report suggests a concrete regression test: send several consecutive short audio clips that cannot be transcribed, then check the behavior on a subsequent normal question and whether any items remain in the provider-side conversation. Tests should also cover normal short utterances to ensure the cleanup mechanism does not delete valid input. The next things to track are a fix commit, test evidence, and inclusion in a release. The available information is insufficient to conclude that all Realtime-compatible providers have the same issue.