Fish Audio Live 3: Full-duplex voice model with 186ms latency

What changed

Fish Audio released Live 3, a full-duplex voice model designed for real-time voice agents. The model is available through the Fish Agent Platform (in public beta), the Fish Audio API, and on-premises deployment for enterprises. Live 3 listens and speaks simultaneously, allowing callers to interrupt, acknowledge, or change direction mid-conversation without breaking the call flow.

The core shift moves conversational turn-taking from hand-tuned pipeline stages into the speech model itself. Traditional voice agents split speech recognition, reasoning, speech generation, and turn-taking across separate components. Fish Audio integrated timing directly into the model, training it on real human conversations to learn when to acknowledge, wait, respond, or stop speaking.

The company demonstrated three specific behaviors the model handles: acknowledging without taking over (saying “okay, 415” while a caller recites a phone number, then waiting), distinguishing between backchannel signals like “yeah” that let the agent continue versus interruptions like “Wait, actually, I just need Thursday” that stop it, and waiting through corrections like “two pizzas, no, three” instead of logging incomplete information.

Fish Audio frames this approach as moving away from “more rules and more thresholds” toward letting the model learn timing from data. A separate reasoning layer still holds the agent’s instructions and tools.

Why it matters

Full-duplex capability addresses a genuine friction point in voice agent design. Most existing systems force users into unnatural turn-taking patterns that require silence or specific keywords to pass control back to the agent. Real human conversations involve overlap, backchannel support, and fluid interruption handling.

For customer service teams, the implications are direct: shorter calls, fewer misunderstandings, and more natural interactions that may reduce customer frustration. For developers, the change simplifies deployment. Rather than tuning silence thresholds and hand-coded turn-taking rules, they can rely on the model to learn timing from training data, reducing engineering overhead for voice agent projects.

The 186 millisecond threshold for short reactions sets a practical reference point. This figure applies specifically to small utterances like confirmations and acknowledgments, not complete responses. Full answers are prepared in parallel. Professionals evaluating voice AI solutions should test whether this latency translates to perceptibly natural conversation flow in their specific use cases, since human perception of responsiveness varies by context and application.

Fish Audio has not published a latency distribution showing how time-to-first-audio performs across different response lengths or load conditions. Third-party evaluation services also have not yet scored Live 3.

What to test

Before adopting Live 3, a thorough evaluation should focus on several areas:

Latency in production scenarios: The 186 millisecond figure covers short reactions only. Test real multi-turn conversations to measure time-to-first-audio for complete responses. Verify how latency scales under load on your deployment option, whether that is the hosted API or on-premises infrastructure.

Turn-taking accuracy: Run scenarios where the model must distinguish acknowledgments from interruptions, mid-sentence corrections, and extended pauses within a single utterance. Determine whether the model makes mistakes that would disrupt calls in ways that hand-tuned rules might not.

Full-duplex versus cascaded modes: Fish Audio supports both full-duplex and cascaded (traditional pipeline) modes. Test both approaches in your specific use case to confirm full-duplex actually improves user experience rather than adding unnecessary complexity.

Language coverage: Live 3 is English-only. The broader Fish Audio platform supports voice generation in 30+ languages, but the full-duplex turn-taking capability is limited to English at launch. Verify whether single-language support fits your requirements.

Benchmark scores and comparisons: Fish Audio publishes no third-party benchmark scores. Establish internal baselines by testing Live 3 against your current voice model or competitors to determine whether the latency and turn-taking improvements justify migration.

Pricing after the promotional period: The $0.035 per minute rate represents 50% off and expires December 31, 2026. After that date, pricing reverts to $0.07 per minute. Factor the full list price into long-term cost projections, not the temporary discount.

Common questions

Can Live 3 handle languages other than English?

Live 3 is English-only. The broader Fish Audio platform generates voice output in 30+ languages, but the full-duplex turn-taking capability of Live 3 specifically is limited to English at this time.

How does the 186 millisecond latency compare to other models?

The 186 millisecond figure applies only to short reactions like confirmations and acknowledgments. Fish Audio has not published benchmark comparisons with other full-duplex or real-time voice models. Third-party evaluation services have not yet assessed Live 3. Direct testing against your current solution or competitors is the most reliable approach.

What does deployment look like for a small team?

New accounts receive a $5 trial balance to start free. The Fish Agent Platform includes browser-based builders and REST API access. For phone support, developers rent a phone number ($1.20 per month) and connect it through the dashboard or API with a single click. The platform handles inbound and outbound calls, web widgets, and integration with external APIs and CRM tools.

The conclusion

Live 3 represents a meaningful architectural shift in how voice agents handle real-time conversation. The full-duplex design and learned turn-taking behavior address genuine friction points in existing voice agent interactions. The 186 millisecond latency for short reactions is competitive, though developers should test end-to-end performance in their own scenarios before committing.

Pricing at $0.035 per minute during the promotional period is reasonable, but the jump to $0.07 after December 31, 2026, aligns with other hosted voice platforms. The primary value lies in reducing engineering complexity around turn-taking rules, rather than in raw speed alone.

The real test comes in production deployments. Whether the model maintains natural turn-taking behavior under load, whether it generalizes well to specific customer service scenarios, and whether the English-only constraint affects your roadmap should all guide adoption decisions. Running a simulation test against your existing voice agent is the most reliable way to evaluate fit.

Related reading

• Mistral Studio: Managed Agents and Enterprise Workflows
• Elephant-Goldfish plugin: spec review for coding agents
• Anthropic Message Threads reduces latency 35% for long conversations