After 186 Milliseconds, Voice Assistants First Need to Learn to Distinguish 'Uh-Huh' Affirmations From Actual User Interruptions

Fish Audio launched Live 3, a full-duplex voice model purpose-built for voice agents. As covered by MarkTechPost author Asif Razzaq, the model can listen while speaking, tell apart casual acknowledgements like "mm-hmm" from deliberate user interruptions, and hits a minimum response time of 186 milliseconds. It only supports English when first released.

Among all its highlighted features, the 186ms response speed is most easily dismissed as a marketing gimmick. But applied to real conversations, this metric has very tangible, noticeable impacts: if the voice agent freezes for half a second right after a user cuts in, the whole user experience drops sharply; if it reacts too quickly and misidentifies every background sound as a user command, the conversation will turn choppy and disjointed. This is exactly the value of full-duplex technology. Most traditional voice assistants use half-duplex design, only starting to listen after they finish speaking, working exactly like a walkie-talkie. Models in the Live 3 category handle listening and speaking simultaneously. The challenge here is not being able to generate speech, but figuring out when to stop talking, when to join the conversation, and when to ignore non-essential sounds completely.

Telling "mm-hmm" affirmations apart from real interruptions is the other side of this same core problem. People constantly make short vocal cues during conversations to signal "I am paying attention". If a voice agent stops speaking every time it hears a short "uh-huh", conversation turns will be cut into tiny awkward fragments; if it does not respond when users actually interject, users will feel like they are talking to a brick wall. This detail sounds trivial, but it is the most common reason users complain about bad voice agent experiences.

The fact that it only supports English at launch also proves that multilingual voice agents are not yet at the stage of being reliably usable. Accents, speaking speeds, acknowledgement habits, and interruption norms all vary widely from language to language. Live 3 has pushed conversation turn management to the center of the voice AI industry's attention. Next, we will see if it can avoid talking over users and ignoring legitimate user inputs in real world dialogues.

Source: MarkTechPost, Asif Razzaq, Fish Audio Releases Live 3, a Full-Duplex Voice Model That Reacts in 186 ms.