Crele Nimbus Tells a Voice Agent Whether the User Is Talking or the TV Is, and Answers Plain-English Questions About the Room Every 40 ms

Voice agents have a familiar failure: they stop talking whenever anyone speaks, including the television. Crele released an open model on October 31, 2026, aimed at that problem. Crele Nimbus listens to a live audio stream and answers yes/no questions about it every 40 milliseconds, starting with two built in: is the user speaking, and is someone else?

The code is on GitHub at Crele-AI/nimbus and the weights on Hugging Face as crele/nimbus-v1, both under Apache-2.0. Crele says the model runs in real time on a laptop CPU and that audio never leaves the machine. The release is very new; the repository and model page showed almost no community activity in their first day.

The problem with ordinary voice activity detection

Voice activity detection, or VAD, is how most voice agents decide that someone has started or stopped talking. Crele’s launch post describes the gap: VAD “hears that someone is speaking, never who,” so an agent “stops for the television, answers a bystander, or waits out someone else’s sentence.”

Nimbus adds speaker awareness. A developer enrolls a few seconds of the user’s voice, and two signals follow the stream: user for that voice, and other for any other voice, whether a person in the room or a loudspeaker. Both can be true at once, which shows when someone talks over the user.

For barge-in, the agent stops only on the user signal’s started event, so a colleague or TV trips other instead and the agent keeps talking. For end of turn, it watches the user stopped event, which another voice does not hold open. By default the voice must stay quiet for 0.3 seconds; Crele notes this detects when the voice stops, “not whether the thought is finished.”

Without enrollment, user answers “can’t tell.” Crele also flags that a voiceprint is personal data in many places and should be stored with consent.

Questions written in plain English

Beyond the two voice signals, developers can ask their own questions, such as “Is music playing?” or “Is someone laughing?”, and add, change or remove them while audio streams. Each question gets its own answer and its own started and stopped events. A session supports up to 32 questions, and an optional yes_means and no_means pair can clarify what a question should count.

Every answer is yes, no or cant_tell, with a probability for yes and a separate probability that the model can judge at all. Four settings, named after common VAD settings, control when events start and stop; the default start threshold is 0.70 held for 0.12 seconds.

What sits under the model

According to the README, about 97 million parameters run on the audio: an 85.5M streaming encoder distilled from Dasheng-Base so that it only hears the past, a 4.5M decision head that reads the last three seconds, and a 7.3M CAM++ speaker-embedding model for the voice signals. An unchanged Qwen3-Embedding-0.6B text encoder reads each question once. Training data includes AudioSet, AudioSet-Strong, LibriSpeech, MUSAN and FSD50K.

Because the audio is encoded once and every question reads from that, Crele says the model keeps up with live audio on an Apple M4 Pro CPU with 8 and with 32 questions. First use downloads about 1.6 GB: 0.4 GB of weights and 1.2 GB for the text encoder.

Crele’s own numbers

The model works on 80 ms chunks and returns two 40 ms answers per chunk, so an answer arrives 80 to 120 ms of audio after the moment it describes, plus compute. On 500 AudioSet clips with default settings, Crele reports a median of 0.24 seconds from a sound starting to a yes, and 0.21 seconds for speech, catching 97% of 181 speech onsets. Questions reworded in ways the model never trained on scored a mean AUROC of 0.892, against 0.911 for its own wording.

The speaker-separation claim is demonstrated on a microphone rather than measured on a benchmark. None of the figures have been independently reproduced.

English only, one voice, tested on a Mac

Questions must be in English, and each session follows one user voice. The first second of every stream answers “can’t tell” while the model gathers context. Input must be 16 kHz mono, and phone audio is untested. CUDA is supported but untested; testing was done on a Mac CPU.

There is no hosted API yet; Crele has opened a waitlist. For voice-agent builders, the release offers a permissively licensed way to make turn-taking depend on whose voice is in the room, not just whether one is.