Sparrow-2 listens to pauses before deciding to respond

Tavus’ new model analyzes speech, pauses, interruptions, and background noise to decide when to listen, speak, or ask someone to repeat themselves.

A pause does not always mean it is time to speak. It may signal a breath, a hesitation, or simply the time someone needs to organize their thoughts. With Sparrow-2, Tavus wants to prevent its conversational agents from responding too soon, remaining silent for too long, or mistaking background noise for someone trying to speak.

Most voice systems rely on a sequence of separate steps: isolate a voice, transcribe it, detect silence, and then determine whether the person has finished speaking. This structure works in a quiet environment, but loses some of the information present in a real conversation.

An “mhm” may encourage someone to continue without being an interruption. A nearby voice may not be addressing the system. A pause lasting several seconds may indicate reflection rather than the end of an answer. By filtering out these signals before interpreting them, an agent risks losing the context it needs to respond appropriately.

Sparrow-2 instead preserves the entire audio scene. The model continuously analyzes the words being used, the meaning of the sentence, rhythm, intonation, speaker identity, backchannels, interruptions, and surrounding sounds. It also keeps track of whether the agent itself is currently speaking.

Based on these signals, Sparrow-2 chooses between several behaviors: continue listening, wait, begin responding, keep speaking, or recognize that the message was not clear enough. In the last case, the agent can ask the person to repeat themselves rather than build a response from an unreliable transcription.

Tavus presents this approach as a response to the “cocktail party problem,” the challenge of following one conversation among other voices and sounds. The company is targeting applications in stores, cars, cafés, shared offices, healthcare facilities, and kiosks installed in busy locations.

The model processes audio as a live stream and updates its estimate of the conversation every 10 milliseconds, or up to 100 times per second. Tavus says it can process 80 milliseconds of audio in around 7 milliseconds, compared with nearly 30 milliseconds for Sparrow-1. That reduction leaves more time for the other components of the agent to prepare and deliver a response.

Sparrow-2 can wait through a six- to eight-second hesitation when it determines that the person has not finished their thought. The previous systems tested by Tavus generally responded after one or two seconds. The model is also designed to distinguish brief listening cues such as “yeah” or “mhm” from a genuine attempt to take the floor.

Tavus trained the model on 16,000 real conversations and 2,000 synthetic conversations that explicitly included interruptions, hesitations, and backchannels. The system uses an audio encoder developed by the company and a six-layer causal model. No timer based solely on the duration of silence determines when it responds.

Tavus reports a conversational failure rate of 2.1%. The figure is based on 575 annotated events taken from 95 production video calls: 95 completed turns and 480 pauses occurring in the middle of someone’s speech. A failure means either interrupting during a pause or never responding after a person has finished speaking.

According to the company’s published results, the strongest alternative it tested made nearly four times as many mistakes. When pauses lasted more than one second, Sparrow-2 allowed 97% of them to pass without interrupting, while Smart Turn and LiveKit v1-mini interrupted approximately half the time.

On a separate set of 96 human-to-human conversations, Sparrow-2 reportedly responded to every completed turn. Smart Turn failed to answer 12.5% of them, while LiveKit v1-mini missed 18%. For the turns that did receive a response, all three systems reached the same median latency of 680 milliseconds.

These comparisons need to be understood within their limits. Tavus tested LiveKit v1-mini, an on-device model, rather than LiveKit’s most advanced turn-taking system. The company chose it to create a comparable setup, but acknowledges that the figures do not represent the full capabilities of its competitor.

The published evaluations also focus exclusively on turn-taking. They do not yet provide the same level of quantitative evidence for Sparrow-2’s ability to understand a noisy room, correctly identify a distant voice, or operate reliably in every environment listed by Tavus. The results were also produced by the company developing the model and have not been independently validated.

Sparrow-2 is now deployed across Tavus PALs, the company’s audiovisual conversational agents, as well as its APIs and platform. A public demo allows users to test how it responds to pauses, interruptions, and background noise.

The change may appear subtle, but it addresses one of the most noticeable flaws in voice assistants: forcing people to change the way they speak. Sparrow-2 attempts to reverse that constraint by asking the system to follow the rhythm of an imperfect conversation, rather than expecting the user to speak as if issuing a command.