What Is Fish Audio Live 3? 186 ms Duplex Voice AI

186 ms Reactions and $0.035 per Minute: A First Look at Fish Audio Live 3

Fish Audio Live 3 is a closed, hosted full-duplex voice AI model developed by Fish Audio that listens while speaking and learns when to acknowledge, wait, respond or stop during real-time conversations. Released on November 10, 2026, it targets phone-calling AI agents and can begin short acknowledgments such as “mm-hmm” in as little as 186 ms, according to Fish Audio, while a separate reasoning layer prepares the full answer. It is available through Fish Agent at a promotional $0.035 per minute, 50% off the regular $0.07 per minute, through December 31, 2026. Live 3’s full-duplex behavior supports English at launch.

Key Takeaways

• Live 3 is a conversational speech model, not another standalone TTS voice. It handles real-time speech input and output while managing conversational timing.
• Full-Duplex Audio means the model keeps listening while it speaks. It can continue after a listener’s “yeah” or stop for a substantive interruption.
• The headline latency is 186 ms for short reactions. It is not a published end-to-end latency measurement for a complete answer.
• Promotional pricing is $0.035 per minute. Telephony adds $0.01 per minute, and premium LLM tokens cost extra.
• Access runs through Fish Agent. The platform is in public beta, while API access is currently limited to select enterprise partners.
• There are no public weights or independent turn-taking benchmarks. Model size, interruption accuracy and latency percentiles remain unpublished.

What Is Fish Audio Live 3?

Fish Audio introduced Live 3 alongside Fish Agent on November 10, 2026, describing both as a move from generating voices toward participating in conversations and completing actions. The official launch announcement positions Live 3 as the conversational model and Fish Agent as the surrounding agent-building platform.

The problem is turn-taking. Conventional agents often combine speech recognition, end-of-utterance detection, LLM reasoning, speech generation and barge-in handling as separate stages. A fast TTS component can still interrupt a caller who merely paused to recall a digit.

Live 3 addresses that orchestration problem inside the speech model. Fish Agent then supplies prompts, voices, knowledge sources, tools, telephony, testing and deployment. Reasoning can still involve a separate LLM, so Live 3 should not be understood as an entire agent stack compressed into one network.

Fish Audio Live 3 at a Glance

Specification	Details
Developer	Fish Audio
Release	Launched November 10, 2026
Type	Closed, hosted full-duplex conversational voice model
Modality	Real-time speech input and speech output
Live 3 language	English at launch
Reaction latency	As little as 186 ms for short acknowledgments; vendor claim
Reasoning	Separate reasoning layer prepares the full response in parallel
Context window	Not yet confirmed
Pricing	$0.035/min promotional; $0.07/min regular
Promotion	50% off through December 31, 2026
Telephony	Additional $0.01/min for inbound or outbound calls
Phone number	$1.20/month dedicated; bring-your-own carrier at no extra cost
Availability	Fish Agent public beta; select enterprise API access; enterprise deployment
License and weights	Proprietary, closed weights; no public download
Web-panel TTS	Not offered as a standalone TTS model

How Fish Audio Live 3 Works and What Makes It Different

Live 3’s distinguishing feature is learned floor control: deciding who should speak, when to remain quiet and when to yield. Five behaviors explain the design.

Full-Duplex Audio keeps the input channel active during speech generation. The agent can therefore interpret what a caller says while its own response is playing.

Learned Turn-Taking replaces some reliance on handwritten silence thresholds. Fish Audio says the model learned when to acknowledge, wait, respond or stop from real human conversations. CEO Rissa Cao compared this learning process with how language models acquire grammar.

Backchannel Acknowledgment separates attentive listening from a request for the floor. During an agent’s response, “yeah” or “mm-hmm” can signal understanding without stopping playback. Live 3 can also insert a quiet acknowledgment while a caller continues speaking.

Interruption Handling covers substantive interventions. If someone says, “Wait, actually, I just need Thursday,” the expected behavior is to stop and listen. The same principle applies when a customer revises an order several times.

Parallel Response Preparation explains the latency claim. A short “okay” or “mm-hmm” can begin in as little as 186 ms while a separate reasoning layer constructs the substantive answer. TestingCatalog’s launch summary describes current reactions as under 200 ms but provides no test conditions.

Live 3’s 186 ms reaction figure applies to short acknowledgments, not to the end-to-end latency of a complete agent response.

Fish Agent also offers Cascaded Mode, which separates more of the conventional voice pipeline. Its builder lists 52 speaking languages and Fish TTS voices using Drama 3 by default or S2.1 Pro. That platform-wide language figure does not extend Live 3’s English-only full-duplex behavior to 52 languages.

What You Can Do With Fish Audio Live 3

• Collect chunked information without premature responses. In Fish Audio’s example, a caller reads “415… 555… 0132,” and the agent can answer “okay, 415” before waiting for the remaining digits.
• Build phone agents that tolerate interruptions. Support, reservation and ordering agents can continue through acknowledgments but yield when callers change their requests.
• Handle mid-conversation corrections. The launch demonstration uses an order changing from two pizzas to three, followed by a request to make one a margherita.
• Give agents knowledge and actions. Fish Agent accepts PDFs, Markdown and URLs as knowledge sources, plus webhook, client and system tools. Builders can also connect an OpenAI-compatible LLM endpoint.
• Deploy across web and telephony channels. Listed options include a web widget, JavaScript and React SDKs, WebRTC, rented or imported Twilio numbers, WhatsApp Business calling and transfer to a human.

An early hands-on post by Charly Wargnier shows a café-server test agent being configured to try these behaviors. It is a useful demonstration of the builder workflow, not independent validation of timing accuracy.

How Fish Audio Live 3 Compares

Live 3 competes on conversational floor control, while ElevenLabs Scribe v3 Live and AssemblyAI-TTS-1 address different stages of the voice stack.

Model	Primary role	Published pricing	Timing or coverage
Fish Audio Live 3	Full-duplex conversational speech	$0.035/min promotional; $0.07/min regular	186 ms short-acknowledgment claim; English
ElevenLabs Scribe v3 Live	Streaming speech-to-text with diarization and turn prediction	$0.22/hour	90+ languages
AssemblyAI-TTS-1	Text-to-speech generation	$15 per 1M input characters	90 ms production first-audio claim; 47 ms lab median time to first byte

These figures are not directly interchangeable. A reaction to a conversational cue, transcription of streaming audio and first generated TTS audio are different measurements. Likewise, per-minute agent pricing, per-hour transcription and per-character synthesis require separate workload calculations.

Availability: How to Access Fish Audio Live 3

Fish Agent is the main entry point. The hosted builder is in public beta, and launch coverage reports a $5 free credit with no card required. Users can configure behavior in plain English, select or clone a voice, attach knowledge and add tools. Launch coverage also lists access to more than 2 million library voices.

API access is narrower than platform access. Fish Audio says it is currently granting API access to select enterprise partners. Interested teams must contact the company rather than assume immediate self-serve API availability. Enterprise deployment, including an on-premises option, is also handled through direct discussions.

Live 3 is not a downloadable checkpoint or web-panel TTS selection. Its distribution is closed and hosted, even when an enterprise negotiates a dedicated deployment.

At the promotional rate, a 10-minute Live 3 phone call costs $0.45 before premium LLM tokens: $0.35 for Live 3 and $0.10 for telephony. The same call costs $0.80 at the regular model rate, excluding any dedicated-number subscription or LLM usage.

Live 3 is not listed in kie.ai’s model catalog. Teams needing only a separate speech-generation component can evaluate Gemini 3.8 Flash TTS, which does not provide Live 3’s duplex turn-taking.

What We Don’t Know Yet

• No independent latency profile is public. Test conditions, concurrency, network contribution and p50, p95 or p99 measurements have not been published.
• Turn-taking quality lacks standardized scores. Fish Audio has not released interruption accuracy, false-barge-in rates or a public turn-taking benchmark.
• Architecture details remain limited. Parameter count, training-data composition, context limits and the precise boundary between Live 3 and its reasoning layer are not yet confirmed.
• Robustness outside demonstrations is unproven. Accents, background noise, compliance scripts and adversarial interruptions need testing on real traffic.
• Broader access and language plans are unclear. A general API release date and support beyond English have not been announced.
• Production adoption is not established. Early public posts consist largely of launch promotion and demonstrations rather than measured customer deployments.

Because weights are not public, there is no local VRAM requirement to assess. For a different deployment problem, an earlier analysis of Nemotron 3.5 Bolt examines a 74B agent model fitting into 24 GB of memory, rather than a hosted speech system.

Frequently Asked Questions

Is Fish Audio Live 3 open source?

No. Fish Audio Live 3 is closed-weight and hosted through Fish Agent; Fish Audio has not published downloadable weights or an open-source license. Enterprise deployment is available by discussion, but that does not make the model open source.

How much does Fish Audio Live 3 cost?

Fish Audio Live 3 costs $0.035 per minute during its 50% promotion, which runs through December 31, 2026, and $0.07 per minute at the regular rate. Inbound and outbound phone calls add $0.01 per minute, a dedicated phone number costs $1.20 per month, and premium LLM usage adds token charges.

How fast is Fish Audio Live 3?

Fish Audio says Live 3 can begin short acknowledgments in as little as 186 ms, while public reports describe reactions under 200 ms. Those figures do not establish the end-to-end latency of a full answer, and no independent latency percentiles have been published.

How do I access Fish Audio Live 3?

You can try Live 3 through Fish Agent, whose platform is in public beta; launch coverage reports a $5 free credit without a card. API access is currently granted to select enterprise partners, and enterprise teams can contact Fish Audio about deployment, including on-premises options.

Is Fish Audio Live 3 a text-to-speech model?

No, not in the conventional standalone sense. Live 3 is a speech-in, speech-out full-duplex model for conversations, and it is not offered as a standalone TTS model in Fish Audio’s web panel; Fish Agent can separately use Drama 3 or S2.1 Pro for cascaded voice output.

What languages does Fish Audio Live 3 support?

Live 3’s full-duplex behavior supports English at launch. The wider Fish Agent builder lists 52 speaking languages for its modes and voices, but that platform figure does not mean Live 3’s duplex timing works in all 52 languages.

Fish Audio Live 3 vs ElevenLabs Scribe v3 Live: what is the difference?

Fish Audio Live 3 is a full-duplex conversational voice model, whereas ElevenLabs Scribe v3 Live is a streaming speech-to-text model. Scribe v3 Live costs $0.22 per hour and supports 90+ languages; Live 3 costs $0.035 per minute on promotion and centers on speaking, listening and turn-taking.

What to watch next. Watch for public API documentation and broader access, independent interruption-accuracy and latency-percentile tests on noisy calls, and confirmation of additional languages. Pricing after December 31, 2026, and evidence of sustained production deployments will also determine whether Live 3 moves beyond convincing launch demonstrations.

Try related models on Kie.ai: Gemini 3.1 Flash TTS and Gemini 3.8 Flash-Lite TTS.