Introduction

Most voice agents fail at the same moment. You stop to remember the middle of your phone number, and they decide you're done. Or you say "yeah" to show you're following, and they stop mid-sentence and apologize. Fish Audio says its new model fixes both.

On November 10 (UTC), Fish Audio launched Live 3, a full-duplex voice model for real-time voice agents. It keeps listening while it speaks, gives small "mm-hmm" and "okay" cues, and has learned to tell an acknowledgment from an interruption. Live 3 is closed and hosted. You can try it today on the Fish Agents platform with a $5 free credit and no card, through the API, or on-premises if you're an enterprise customer.

What Shipped

  • Model: Live 3, Fish Audio's full-duplex real-time voice model, announced in the launch blog and on
    .
  • Modality: speech in, speech out for voice agents, with timing built into the speech model itself instead of rules spread across the pipeline.
  • Language: English only for now.
  • Access: closed weights. Use it on the Fish Agent Platform (in public beta), through the API, or on-premises for enterprises.
  • Price: $0.035 per minute, shown as 50% off a $0.07 list price, with no end date posted. Phone calls cost $0.01 per minute extra, a phone number costs $1.20 per month, and premium LLMs are billed at their own token rates.
  • Voices: any of Fish Audio's 2M+ library voices, or your own clone.

What It Actually Does

Fish's co-founders, Shijia Liao (chief scientist), Jiahua Liu (CTO) and Rissa Cao (CEO), say Live 3 was trained on real human conversations. It learned when to acknowledge, wait, respond or stop from those calls, instead of following hand-written thresholds. "The fix so far has been more rules and more thresholds," Cao said. "We think a model should learn this from real conversations, the same way language models learn grammar."

The blog walks through three behaviors:

  • Backchannels during chunked input: read out "415… 555… 0132" and Live 3 may say "mm-hmm" or "okay, 415" between the chunks, then wait for the rest.
  • Acknowledgment vs. interruption: while it lists services or reads an order back, "yeah" and "mm-hmm" let it keep going. "Wait, actually, I just need Thursday" makes it stop and listen.
  • Self-corrections: "Two pizzas… no, three. Three large. And… actually, make one a margherita." It waits through each change instead of confirming the wrong order.

The one hard number: Live 3 can decide to respond in as little as 186 milliseconds. Quick replies like "mm-hmm" keep the exchange going while the full answer is still being prepared. Fish hasn't published turn-taking benchmarks, interruption-accuracy scores or end-to-end latency percentiles, so judge it by the sample calls in the blog and by calling it yourself.

The Platform Around It

Live 3 ships inside Fish Agents, a hosted builder that's in public beta for every Fish Audio account:

  • Build: a system prompt, first message, voice, speaking language, knowledge base (PDFs, Markdown, URLs) and tools (webhooks you can import from cURL, client tools and system tools). You can also bring your own OpenAI-compatible LLM endpoint.
  • Modes: full-duplex or cascaded. The builder lists 52 speaking languages and voices agents with Fish's TTS (Drama 3 by default, or S2.1 Pro), but Live 3's full-duplex behavior is English only for now.
  • Deploy: a two-line web widget, @fishaudio/agent-client and @fishaudio/agent-react SDKs, WebRTC, rented or imported phone numbers (Twilio), WhatsApp Business calling, and warm or cold transfer to a human.
  • Test and monitor: LLM-judged single-turn, multi-turn and tool-call tests, version history, recordings, turn-by-turn transcripts, a latency breakdown (end-of-utterance, LLM and speech), and post-call summaries sent by webhook.
  • Privacy: zero data retention is offered.

To get people talking, Fish's team replied to launch-day requests on X with free custom agents, each built in about 5 to 9 seconds. They included a printer helper, a contractor job-intake line and a "brutally honest rubber duck" debugging partner.

How to Try It

The fastest way is the browser. Open fish.audio/agents, start from a template or a one-line description, and make a preview call. To manage agents in code, use the REST API under /v1/agent/*:

curl https://api.fish.audio/v1/agent/agents \
  -H "Authorization: Bearer $FISH_API_KEY"

Then bind a phone number to an agent:

curl -X POST https://api.fish.audio/v1/agent/phone-numbers \
  -H "Authorization: Bearer $FISH_API_KEY" \
  -d '{"provider": "twilio", "phone_number": "+14155550123", "agent_id": "agt_your_agent_id"}'

Pro Tip

Test Live 3 the way real callers break agents. Read a 10-digit number in three chunks with long pauses. Say "uh-huh" twice while it's reading something back. Then cut in with a correction. If it waits, keeps going and then yields at the right moments, the full-duplex part is working. Keep your test agent's speaking language set to English, because the full-duplex model doesn't cover other languages yet.

Where It Fits

Live 3 is for phone support, bookings, restaurant orders, intake lines and reminder calls, where people hesitate, read out numbers and change their minds. It's not an open model. If you want full-duplex weights to run yourself, look at Sophon-Realtime-v1. If you want a hosted persona model, compare it with Nuance Labs' Emma / Nuplex. Inside Fish's own stack, Live 3 builds on the turn-taking work in its transcribe-stream-1 preview and the directable voices of Drama 3.

Conclusion

Live 3 isn't about a prettier voice. It's about timing: knowing when to say "mm-hmm", when to keep talking and when to stop. Fish has put that timing into the speech model itself, priced it at three and a half cents a minute for launch, and made it available to any account today, in English only. There's one hard number (186 ms to decide) and no third-party benchmarks yet, so make a call, read out your phone number slowly, and see if it waits for you.

—Anabel ♡