Fish Audio releases full-duplex Live 3 voice model

Fish Audio has released Live 3, a full-duplex voice model built for voice agents, according to MarkTechPost. The release matters because it targets one of the hardest parts of conversational AI: allowing a system to receive audio while it is also producing speech, without treating every sound as a stop signal.

What did Fish Audio release?

According to MarkTechPost, Fish Audio released Live 3 as a full-duplex voice model for voice agents. The report says the model can keep listening while it is speaking, which is the defining behavior of a full-duplex system. It also says Live 3 can tell the difference between a brief listener acknowledgment such as “mm-hmm” and a genuine interruption. At launch, the model is described as English only. Those details make Live 3 a focused release for teams building conversational products, rather than a broad multimodal platform update. The release is narrow but meaningful: it is not presented as a general chat model or a media generator, but as infrastructure for real-time conversation. For builders, that focus can be more useful than a broad feature list because it targets a specific failure mode in voice products.

What does full-duplex mean for voice agents?

In voice technology, full-duplex generally means that both sides can send and receive audio at the same time. That is different from a simpler turn-taking setup, where the system waits for the user to stop before it responds. For voice agents, the difference is practical. Human conversation often includes overlapping sounds, quick acknowledgments, and mid-sentence corrections. A model that can continue receiving input while producing speech has more information to decide whether the user is agreeing, thinking, or actually taking over. MarkTechPost frames Live 3 around that use case, which suggests Fish Audio is aiming at agents that need to feel more like live participants than request-and-response tools. This also changes how designers should think about silence. In a turn-taking system, silence is often treated as the end of a turn. In a full-duplex system, silence can be part of the interaction, and the agent must decide whether to wait, prompt, or continue.

How does Live 3 separate “mm-hmm” from real interruptions?

The most specific behavior in the report is Live 3’s ability to distinguish a backchannel from an interruption. A backchannel is a short sound or phrase that shows the listener is still engaged, such as “mm-hmm,” “right,” or “okay.” An interruption is different: the speaker is trying to take control of the conversation, change the topic, or stop the agent from continuing. If a voice agent treats every sound as a stop signal, it may cut itself off too often. If it ignores every sound, it may miss a real request. According to MarkTechPost, Live 3 is designed to make that distinction, which is a useful capability for natural-sounding voice agents.

Why does a 186 ms reaction time matter?

MarkTechPost says Live 3 can react in as little as 186 milliseconds. That number is best read as a best-case latency figure, not a promise that every interaction will feel that fast. Still, reaction time is central to voice-agent quality. When a system responds too slowly, users may wonder whether it heard them. When it responds too quickly, it may interrupt before the user has finished. A fast full-duplex model gives designers more room to tune those tradeoffs. For developers, the reported figure is a benchmark to test under real conditions, including network delay, microphone quality, and the agent’s surrounding workflow. A reported latency figure is most useful when paired with a test plan. Teams should measure perceived responsiveness in realistic conversations, not only in isolated benchmark runs.

What are the launch limits?

The report’s clearest limitation is language. MarkTechPost says Live 3 is English only at launch. That matters for product teams because voice agents are often expected to support multiple languages or regional accents. An English-only launch can still be useful for demos, internal tools, or English-first services, but it narrows the initial market. The report also does not provide a full deployment picture, such as pricing, regional availability, or integration details. Readers should treat those as open questions until Fish Audio publishes more information. For now, the safe summary is that Live 3 is a specialized voice-model release with a clear interaction focus and a stated English-only launch scope. That does not make the release unimportant. It simply means the first audience is likely to be English-language voice-agent developers who need better turn-taking behavior.

How does this relate to Grafy?

Based on the provided Grafy app list, there is no direct Live 3 integration and no real-time full-duplex voice runtime. That means readers cannot simply open a Grafy app and run Live 3 as a live voice agent based on the available information. The connection is still useful for teams thinking through the surrounding workflow. Grafy Bots can help design embeddable support-bot behavior, including dialogue rules and escalation paths, before a voice layer is added. Grafy Chat can be used to compare how different models handle conversational prompts and clarification strategies. Grafy Reader can help teams read and question papers or reports about voice-agent design. From the Grafy News Desk, the practical approach is to treat Live 3 as a signal about where voice agents are heading, then use Grafy’s existing tools to prepare the product logic around that future voice layer.

Frequently asked questions

Does Live 3 listen while it speaks?

Yes. According to MarkTechPost, Live 3 is a full-duplex voice model, meaning it can keep listening while it is speaking. That design is aimed at voice agents that need to behave more like a live conversation, rather than waiting for a speaker to finish before capturing the next input.

How quickly can Live 3 react?

According to MarkTechPost, Live 3 can react in as little as 186 milliseconds. That figure describes the fastest reported response time, not a guaranteed latency for every deployment. For voice agents, lower reaction times can make pauses feel shorter, but real-world performance would still depend on network conditions, audio quality, and the surrounding system.

Is Live 3 English-only at launch?

Yes. MarkTechPost says Live 3 is English only at launch. That means the initial release is described as limited to English, according to the report. Teams planning multilingual voice agents should treat that limitation as a launch constraint and check for later language support before relying on the model for other languages.

---

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