Fish Audio Launches Fish Agent Beta on Its Live 3 Full-Duplex Speech Model

Fish Audio has released Fish Agent, a no-code platform in public beta for building AI voice agents that can place and answer phone calls. The service runs on Live 3, the company's new full-duplex speech model designed for real-time conversation, and Fish says a developer can describe an agent, attach a phone number and publish it in roughly 30 seconds.

Live 3 targets a persistent weakness in voice applications: deciding when to speak. Fish reports the model can choose a brief reaction such as "mm-hmm" or "okay, 415" in as little as 186 milliseconds. Because the model is full-duplex, it keeps processing incoming speech while generating audio, so it can acknowledge, wait or stop as a conversation unfolds.

Turn-taking handled inside the model

Conventional voice agents usually chain automatic speech recognition, a large language model and text-to-speech, with a voice activity detector estimating from volume and silence when the caller has finished. A short thinking pause can set off a premature reply, while a brief acknowledgement like "uh-huh" can be read as a new turn.

Live 3 instead places those floor-control decisions inside a speech model trained on human conversations. According to Fish, it can listen while it speaks and choose among acknowledging the caller, continuing its response, yielding the floor or waiting.

From prompt to phone line

Fish Agent bundles Live 3 with the pieces needed to deploy a working application. Developers set a system prompt, select or clone a voice, attach reference material, connect tools and then publish to the web, a phone number or an application.

• Voices: more than 2 million catalog voices across 30-plus languages, custom voice clones, adjustable speed, an expressive mode and three model tiers. The capability list states that Live 3 supports 30+ languages at launch.
• Knowledge: grounding from uploaded PDFs, Markdown files and URLs.
• Tools: API endpoint imports from cURL for in-call actions such as refunds, bookings and CRM updates.
• Telephony: numbers rented through the dashboard or connected through an existing carrier.
• Testing: single-turn, multi-turn and tool-call tests scored by an LLM against success examples supplied by the developer.
• Observability: recordings, transcripts, call summaries and latency breakdowns for end-of-utterance detection, LLM generation and speech synthesis.

Assigning telephony over REST

Phone-number assignment uses the REST endpoint at /v1/agent/phone-numbers. The request below connects a Twilio number to an existing agent:

curl -X POST https://api.fish.audio/v1/agent/phone-numbers \
  -H "Authorization: Bearer $FISH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "twilio",
    "phone_number": "+14155550123",
    "agent_id": "agt_a1b2c3d4e5f6"
  }'

Launch pricing under four cents

Fish lists the combined speech-recognition, LLM and voice bundle at $0.035 per minute during its launch promotion. The standard rate is $0.07 per minute, with telephony and number rental billed separately.

The API is hosted at api.fish.audio under /v1/agent/*. Fish has stated that the promotional rate ends on 31.12.2026.

Context around the 186 ms claim

The 186 ms figure measures the model's decision to produce a short reactive utterance. The announcement provides no time to first audio for a complete answer, no tail latency under load and no interruption accuracy, all of which directly affect whether an agent talks over callers or leaves long gaps.

Fish has also published no benchmark scores for Live 3, and at launch Artificial Analysis has no Live 3 entry in its duplex-speech harness, so the available performance numbers remain vendor-reported. The piece recommends that a production evaluation measure:

• Median and 95th-percentile time to first audio for complete responses.
• False interruptions during pauses, background noise and brief acknowledgements.
• How quickly and reliably the agent stops when the caller takes the floor.
• Tool-call accuracy, task completion and recovery from API failures.
• Audio quality over real phone networks, including noisy and low-bandwidth calls.

One stack, one runtime

Fish Agent competes with OpenAI's Realtime API, ElevenLabs' Realtime v1, Google's Gemini Live and agent platforms built around LiveKit or Vapi. Fish supplies the duplex model, agent runtime, telephony, testing and observability as one hosted service, and it also advertises an on-premises version of the integrated stack for enterprise customers.

Live 3 has no standalone real-time API at launch: access runs through Fish Agent, so developers who choose the model also adopt Fish's runtime and deployment model. A modular architecture can still pair separate speech-recognition and LLM services with Fish Audio text-to-speech, or place Fish components inside another agent framework.

Production buyers will need to confirm data-retention policies, deployment regions, compliance certifications, concurrency limits, rate limits and service-level commitments; the public launch details do not specify those operational terms.

Where the beta fits

Fish Agent's initial feature set fits voice support, outbound sales, scheduling and reminder systems operating in English. Its knowledge grounding, tool calls, telephony and pre-release tests cover the main components of a transactional voice workflow without separate vendors for each stage.

Applications that require multilingual conversation cannot use Live 3 for that need at launch, despite the broader language coverage in Fish Audio's voice catalog. Teams that need independently verified latency or quality scores likewise lack enough public evidence for a benchmark-based vendor decision.

Early pilots should compare Live 3 with the existing production stack using recorded call scenarios, noisy phone audio, caller interruptions and real tool failures. Those tests will show whether its learned turn-taking improves completed tasks and call flow beyond Fish Audio's controlled demonstrations.