On November 10, 2026, Fish Audio announced Live 3, a full-duplex voice model. It keeps listening to the other person while it speaks, and the model itself handles the rhythm of conversation: backchannels, interruptions, and mid-sentence corrections.

Live 3 - fish audio

First, the most important point: Live 3 currently supports English only. For Japanese and other non-English languages, speech is automatically generated by a separate voice engine, Fish TTS (the Drama 3 model).

This article covers what Live 3 can do, its limitations, how to approach non-English use, and pricing. It’s based on the official announcement page, the product page, and answers from official support.

The Short Version: What Can Live 3 Do, and What About Other Languages?

  • Live 3 is a “full-duplex” voice model that listens while it speaks.
  • Its three signature conversational behaviors are giving backchannels, stopping only for real interruptions, and waiting while you correct yourself.
  • English is the only supported language. For other languages, it automatically switches to Fish TTS.
  • Pricing is $0.035 per minute (50% off the regular $0.07, for a limited time through December 31, 2026).
  • It’s a strong option if you want to build phone or support agents for English-speaking audiences. If you want to build a voice agent in another language, you can’t take advantage of what makes Live 3 special.

What Is Live 3?

Live 3 - fish audio

Live 3 is a full-duplex voice model built into Fish Audio Voice Agents, Fish Audio’s voice agent platform.

Typical voice agents use a pipeline that combines speech recognition, reasoning, speech generation, and turn-taking (controlling who speaks when) as separate components. Making them work together naturally takes a lot of fine-tuning.

Live 3 builds this “conversational timing” directly into the voice model. According to Fish Audio, the model was trained on real human-to-human conversations and has learned when to give a backchannel, when to wait, when to respond, and when to stop talking.

It listens while it speaks, and it has a separate reasoning layer for agent instructions and tool execution. In other words, you still decide what to say (instructions and tools) through the system prompt, while Live 3 manages when and how to say it.

Live 3’s Three Conversational Behaviors

1. Backchannels Without Taking Over

Imagine someone reading out a phone number in chunks: “415… 555… 0132.” Live 3 drops in short responses like “Mm-hmm” or “Got it, 415” between chunks, then waits for the remaining digits.

These backchannels do two things: they confirm that the digits came through, and they leave the floor open for the user to keep talking.

Live 3 can also decide how to respond in as little as 186 milliseconds, according to Fish Audio. Short reactions like “Mm-hmm” or “Got it” and quick confirmations keep the conversation moving while the full answer is still being prepared. Note that 186 ms is the best case, not an average response time.

2. Stopping Only for Real Interruptions

Not every sound is an interruption. If the agent is explaining business hours and the user says “yeah” or “uh-huh,” Live 3 keeps going.

But if the user says something like “Oh, wait. Actually, I only need Thursday,” it stops talking right away and makes room for the new request.

This distinction means users can give backchannels without worrying about cutting the agent off, and they don’t have to wait for the end of a long answer when they want to change direction.

3. Following Along as You Correct Yourself

“Two pizzas… no, three. Three larges. And… actually, make one of them a Margherita.”

In the middle of a correction like this, Live 3 doesn’t lock in the wrong order. It waits for the correction to finish. It understands that the user is still deciding, and gives them room to settle on what they want. In support, booking, and ordering scenarios, hesitations and corrections like this happen all the time.

Phone order - fish audio

Live 3’s Limitations

Keep the following limitations in mind before using Live 3.

ItemHow Live 3 handles it
Supported languagesEnglish only
Other languagesAutomatically switches to Fish TTS (Drama 3 model)
Speech speed adjustmentNot supported
Expressive modeNot supported
Pronunciation dictionaryNot supported
Speech recognitionFixed to Fish Transcribe Stream 1

Note that some settings you can adjust with Fish Audio’s standard voice features aren’t available in Live 3.

Using Live 3 in Languages Other Than English

Live 3 is English-only for now, so agents in other languages automatically generate speech with Fish TTS. According to Fish Audio support, to change the agent’s spoken language to something other than English, you switch the engine over to Fish TTS in the settings.

Once you switch to Fish TTS:

  • The agent can speak other languages. Fish Audio states support for 2 million+ voices and 30+ languages.
  • The full-duplex behaviors that are Live 3’s main selling point (backchannels, interruption detection, and following corrections) are described as Live 3 features, so it’s safer not to expect the same experience in other languages.
  • Platform features such as the knowledge base, tool integrations, testing, and call analytics remain available.

Pricing

Here is Live 3’s pricing as listed on the product page.

  • Call rate: $0.035 / minute (50% off the regular $0.07 / minute, for a limited time through December 31, 2026)
  • Phone calls (inbound and outbound via a phone number): +$0.01 / minute
  • Phone number: $1.20 / month (no extra charge if you bring your own carrier line)
  • Premium LLM (when switching to a larger model for complex reasoning): billed separately at that model’s token rates
  • You start with $5 in credit, and no credit card is required

For example, here’s a rough estimate for renting one phone number and handling 100 minutes of phone calls per month (calculated at the limited-time price, with no extra LLM charges):

ItemCalculationAmount
Call rate$0.035 × 100 min$3.50
Phone calls$0.01 × 100 min$1.00
Phone numberMonthly fee$1.20
Total$5.70

Who It’s For and Who It’s Not For

Pros 5
  • Well suited to anyone building phone, support, booking, or ordering agents for English-speaking audiences

  • Good for use cases where natural backchannels and interruption handling directly affect the quality of the experience

  • Comes with $5 in starting credit, so you can try it without entering a credit card

  • Includes production-ready features (knowledge base, tool integrations, testing, and call analytics)

  • Offers zero data retention and on-premises deployment for enterprise customers

Cons 4
  • English only. Other languages switch to Fish TTS, and Live 3's conversational behaviors aren't available

  • Speech speed, Expressive mode, and the pronunciation dictionary don't work

  • The 50% discount is limited to a period ending December 31, 2026

  • For voice agents in non-English languages, you can't take advantage of Live 3's strengths

How to Get Started

  1. Sign up for Fish Audio

Create an account on the official website. You get $5 in credit, and no credit card is required.

  1. Create an agent

Choose a template, then set up the system prompt (personality, goals, guardrails) and the first message.

Agent settings - fish audio

  1. Set up knowledge and tools

Add PDFs, Markdown files, or URLs to the knowledge base. If needed, import endpoints from cURL so the agent can call your own APIs.

  1. Run a test call in your browser

Check the agent’s behavior with a test call before going live. Testing tools are available, from one-off checks to full call simulations.

  1. Go live

Rent a phone number, or connect through your website, the Web SDK, or WebRTC to publish.

Summary

Live 3 is a full-duplex voice model that handles the rhythm of conversation (backchannels, interruptions, and corrections) in the model itself. For anyone building phone agents for English-speaking audiences, it’s a new model worth a close look.

That said, it currently supports English only. Voice agents in other languages use Fish TTS for speech generation, which means Live 3’s conversational behaviors aren’t available. I couldn’t find any plans for additional language support in the official information reviewed here, so check Fish Audio’s official announcements for updates.