TTS API for Voice Agents: How to Choose One (2026)

The best TTS API for a voice agent is the one that passes your real call tests without weakening voice identity, latency or reliability. For most existing LiveKit or Pipecat teams, Fish Audio is the strongest fit when expressive delivery and reusable voices are the hard requirement; Cartesia is the latency-first alternative, ElevenLabs fits broader studio workflows, and OpenAI fits teams already standardized on its APIs and preset voices.

That is a segmented recommendation, not a universal ranking. The provider is only one stage in a speech loop that also includes telephony, speech recognition, an LLM, tools, context, interruption and fallback.

Use the AI voice agent guide to design that complete loop before comparing provider contracts.

What should a voice-agent TTS API support?

A production TTS API needs streaming audio, a stable voice identifier, low time to first audio, interruption-friendly output, multilingual pronunciation and predictable concurrency. It should also expose enough controls to keep a series or brand voice consistent.

The orchestration layer remains responsible for barge-in, turn detection, tool calls, context and human escalation. A provider that sounds excellent in a sample but cannot fit the streaming contract still fails as an agent component.

How do the main TTS providers compare?

Fish Audio is built around expressive speech, voice cloning from short samples, a large community voice library and one-time credit packs. It fits teams that need a recognisable agent or creator voice and can own the surrounding orchestration.

Cartesia positions its Sonic models around low-latency real-time speech and voice-agent use cases. It is a strong shortlist when the first audible response is the dominant constraint and its voice set fits the brand.

ElevenLabs offers a mature voice and studio ecosystem with broad language support and established integrations. It fits teams that value its voice catalogue, editing tools or wider content-production surface beyond conversational agents.

OpenAI TTS fits applications already committed to OpenAI models and SDKs. Its preset-voice workflow can reduce integration choices, but teams that need reusable custom identity should verify the current voice and cloning capabilities rather than assuming parity.

LiveKit Inference is not a replacement for every direct API. It is a managed layer that can expose several providers, including Fish Audio, inside LiveKit's infrastructure. Choose it when one cloud vendor and no separate Fish Audio key outweigh the control of direct billing.

Verify every current capability and price in the provider's own documentation. This page is a decision framework, not a claim that product contracts will remain static.

How should you compare latency?

Measure from the moment the caller stops speaking to the first audible agent syllable. Break the result into STT finalization, LLM first token, TTS time to first audio, network transport, jitter buffer and playback.

A useful voice agent often begins speaking within one second, with many production systems targeting a total p95 below 1.5 seconds. The right threshold depends on the interaction. A short acknowledgement can arrive much sooner than a policy answer that requires a retrieval step.

Measure p50 and p95 across real scripts. Do not compare an API's isolated inference time with a competitor's complete call recording. The real-time text to speech guide explains the complete latency budget.

How should you compare price?

Convert every option to the cost of the speech your agent actually produces. Include greetings, confirmations, retries, repeated explanations and portions generated before a caller interrupts.

At roughly 140–150 English words per minute, a 10-minute spoken floor is about 9,000–10,000 characters. If an agent repeats confirmations or regenerates failed chunks, the billable volume can exceed the final call transcript.

Fish Audio on fishaudio.lol uses one credit for about 1,000 characters and includes API access in the Studio pack. Use the TTS API pricing guide for a worked character-to-credit estimate, then add the agent platform, telephony, STT, LLM and fallback provider fees.

Which provider fits which voice agent?

Choose Fish Audio when the voice itself is part of the product and the team needs expression, cloning or a reusable identity. The direct Vapi, Retell, LiveKit and Pipecat guides explain the integration boundary.

Choose Cartesia when the shortlist is driven by managed latency and its voice set already satisfies the brand. Review Cartesia's current pricing and language support before basing a budget on a comparison article.

Choose ElevenLabs when the team needs its larger creation and studio ecosystem in addition to speech generation. Use the Fish Audio vs ElevenLabs comparison for a focused cost and voice workflow decision.

Choose OpenAI when architecture simplicity inside an OpenAI application matters more than reusable custom identity. Check the current OpenAI speech pricing and voice contract before launch.

How should you run a production pilot?

Create one script set that covers a greeting, a short answer, a long tool-driven answer, an interruption, a clarification and an escalation. Run the same scripts through every shortlisted provider and agent framework.

Include product names, prices, dates, account identifiers and two languages. Test fallback by forcing a provider timeout or invalid voice. Confirm that the caller hears either a sensible alternate voice or a controlled transfer, not silence.

Review p95 latency, failed generations, voice drift, pronunciation corrections and cost per completed call. Keep the recording, model ID, voice ID and configuration version with the decision.

What security and consent checks belong before launch?

Keep provider keys server-side and scope them to the smallest production environment that can operate the agent. Do not expose Fish Audio, Cartesia, ElevenLabs or OpenAI credentials through public runtime configuration.

For cloned voices, retain documented permission for the caller experience, languages, channels and duration. Test revocation, deletion and fallback. A cloned voice needs a different permission review from a licensed stock voice.

Log provider latency and failure classes without retaining unnecessary caller content. Treat transcripts and recordings as customer data with explicit retention rules.

Key takeaways

• Start with the agent framework, latency target and voice-rights requirements.
• Fish Audio fits expressive identity and cloning; compare Cartesia, ElevenLabs, OpenAI and LiveKit Inference by actual need.
• Measure p95 from the caller's last syllable to the first audible reply.
• Budget characters spoken plus retries, not just final transcript minutes.
• Keep voice IDs and provider configuration portable across agent versions.

Frequently asked questions

What is the best TTS API for voice agents?

There is no single winner. Fish Audio fits teams that need expressive, reusable voice identity; Cartesia fits latency-first managed agents; ElevenLabs fits teams needing its broader voice and studio ecosystem; OpenAI fits teams already standardized on its APIs and preset voices.

Which latency metric matters most for a voice agent?

Measure p95 time from the caller finishing a turn to the first audible reply. Full conversational latency includes STT finalization, LLM generation, TTS time to first audio, transport and playback.

Does a TTS API include interruption handling?

No. The agent framework must cancel in-flight speech, preserve the transcript and decide how the next turn continues. The TTS provider generates speech; it does not own the complete barge-in policy.

How should I compare voice-agent TTS pricing?

Convert each provider to the cost of spoken characters or minutes at your monthly volume, then add retries, agent-platform fees, telephony, fallback and concurrency. Compare the approved voice class rather than the cheapest listed model.

Can I switch TTS providers after launch?

Yes if voice IDs, audio formats and orchestration are isolated behind your own configuration. Keep scripts, pronunciation rules, consent records and call tests independent from one provider contract.