Retell AI Text to Speech: Voice Agent Setup Fish Audio
Evaluate Fish Audio voices for Retell AI agents: provider compatibility, voice cloning, latency tests, fallback, rights and per-minute cost planning.
Retell AI text to speech should be evaluated as a provider and latency decision: which Fish Audio voice runs the agent, how the provider is configured, what Retell supports today and what happens when speech generation fails. Fish Audio publishes a Retell AI partner workflow, but fishaudio.lol should not claim a one-click integration until the current account-level configuration confirms it.
The useful test is a real Retell call with a real script, not a voice sample played on a product page.
What does Retell handle versus Fish Audio?
Retell manages the agent conversation, telephony, turn-taking, tools and call analytics. Fish Audio supplies expressive text to speech and reusable voice identity.
Your configuration needs to select a supported voice provider, authenticate to Fish Audio through the approved path and return audio in Retell’s expected format. Verify the current provider list in Retell’s docs rather than relying on an old tutorial.
How do you choose the agent voice?
Pick the voice from a real support or sales response, not a dramatic character line. Test:
- A calm explanation of a delay.
- A short confirmation of an order number.
- A question that requires a rising intonation.
- An interruption and correction.
- A transfer to a human.
Fish Audio’s voice library can help shortlist candidates. For a brand voice, read the voice cloning API workflow and collect the speaker consent before configuration.
How should Fish Audio be configured?
Keep the Fish Audio key on the backend or provider integration, never in a browser or public agent configuration. Store the selected voice ID with the Retell agent version.
Use a staging key and a staging Retell agent first. Confirm the provider request shape and audio response against the current contracts before switching the main phone number.
If the native path cannot represent the required voice or response format, use a bridge you operate and document every field translation.
What latency targets matter?
Aim for less than one second from the end of the caller’s turn to the first audible reply, then track p95. Break latency into Retell transcription, agent/model work, retrieval, Fish Audio time to first audio and telephony playback.
The real-time text to speech guide covers chunking, warm connections and cancellation.
How should fallback and transfer work?
Configure a fallback voice or provider when Fish Audio times out, runs out of credits or rejects the voice. Return explicit errors; an empty audio success can strand the caller.
For high-risk requests, transfer to a person with a short summary and the approved transcript. A voice provider must not become the blocker between a caller and help.
How much does the speech cost?
At roughly 150 words per minute, one minute of speech is about 900 characters or one Fish Audio credit. API access starts with the $99 Studio pack and 15,000 credits.
| Use | Fish Audio cost | Separate cost |
|---|---|---|
| One minute of agent speech | About 1 credit | Retell, telephony |
| 30-minute support day | About 30 credits | Retell, telephony |
| 1,000 call minutes | About 1,000 credits | Retell, telephony |
The TTS API pricing guide converts character volume into a budget.
Which compliance controls belong in the pilot?
Disclose automation and recording where required, minimize retained audio and protect transcripts as customer data. For cloned voices, keep consent, project scope and revocation instructions with the voice configuration.
Test that a caller can request a human and that the transfer carries the context needed to continue without repeating the whole story. Review provider logs to confirm keys, voice identifiers and customer text are not exposed beyond their intended systems.
The AI voice agent guide covers the broader rollout, while the commercial use guide explains Fish Audio pack rights.
What should the pilot review?
Review answer quality, transfer rate, interruption behavior, p95 latency, failed generation and voice consistency. Keep a fixed script and a set of real call scenarios so changes can be compared.
Do not mark the integration complete because a test sample played successfully. A production agent needs a provider fallback path and a human escalation policy.
Key takeaways
- Retell owns the agent and telephony; Fish Audio supplies speech.
- Verify the current native/provider path before claiming one-click support.
- Keep voice IDs, keys and consent records under version control.
- Measure p95 latency and generation fallback on real calls.
- Budget roughly one Fish Audio credit per minute of speech.
Frequently asked questions
Is Fish Audio available inside Retell AI?
Fish Audio publishes a Retell AI partner workflow. Verify the current Retell provider list and the exact configuration available to your account before promising a one-click setup on fishaudio.lol.
Can Retell agents use a cloned Fish Audio voice?
A cloned voice can be an objective if the current provider path returns a reusable Fish Audio voice identifier. Confirm compatibility and keep documented speaker consent.
What latency should I measure?
Measure from the caller’s last syllable to the first audible reply through the real Retell call. Track p95 rather than only the average.
What happens if the custom voice provider fails?
Configure a supported fallback provider or safe handoff in Retell. Your bridge should return a clear error when Fish Audio times out, lacks credits or receives an invalid voice.
How is Fish Audio priced for a Retell agent?
Fish Audio consumes one credit for about 1,000 generated characters; API access comes with the $99 Studio pack. Retell and telephony charges are separate.