Vapi Text to Speech: Custom Voice Integration Fish Audio

Vapi text to speech can use a custom speech provider when the provider contract, audio format and latency fit the assistant. Fish Audio can supply expressive TTS and cloned voice identity through a backend bridge; Vapi continues to orchestrate turns, tools, telephony and fallback.

This is a provider-integration decision, not a claim that every Fish Audio feature appears as a native Vapi option.

How should Vapi and Fish Audio divide the stack?

Vapi owns the assistant flow: call control, speech-to-text selection, model/tool execution, turn-taking and provider fallback. Fish Audio supplies text-to-speech generation and reusable voice identity.

Your integration must map Vapi’s current request into a Fish Audio request, return audio in the expected shape and report failures quickly enough for Vapi to apply a fallback.

Review Vapi’s custom TTS documentation and voice provider overview before implementation. Contracts can change; the live docs are authoritative.

What makes a custom TTS provider compatible?

The provider must accept the text and voice parameters Vapi sends, return the supported streaming or file response and complete within the assistant’s latency budget.

Check:

• Authentication and server-to-server request shape.
• Audio encoding, sample rate and channel layout.
• Streaming behavior and end-of-speech signaling.
• Timeout, retry and non-2xx error propagation.
• Voice identifier mapping and fallback voice.

If any field cannot be represented cleanly, use a small bridge service you own. Do not silently reinterpret one provider’s contract as another’s.

How do you keep a branded voice consistent?

Choose one Fish Audio voice per assistant and store its identifier beside the Vapi assistant version. If the voice is cloned, retain the consent record and use the voice cloning API workflow as the source of voice lifecycle policy.

When a voice changes, version the assistant. Historical calls should remain attributable to the old voice configuration; new calls use the approved replacement.

How should latency be measured?

Measure from the end of the caller’s speech to the first audible reply on the real Vapi path. Break the result into transcription, model, tool, custom TTS and playback.

Set a p95 budget and a fallback behavior:

• Retry a transient TTS failure within a small time limit.
• Switch to a supported Vapi provider if the custom service misses the budget.
• Cancel generation when the caller interrupts.
• Stop retrying authentication or malformed-request errors.

The real-time text to speech guide explains why average latency hides bad conversations.

How do you handle failure and fallback?

Your bridge should preserve the upstream status class and distinguish timeout, rate limit, invalid voice, exhausted credits and malformed text.

Do not return an empty success response. Vapi needs a clear error to trigger the configured fallback. Keep the original text, voice ID and request ID in server logs without exposing keys or customer conversations.

How much does the speech cost?

Fish Audio API access starts with the $99 Studio pack: 15,000 credits, roughly 15 million generated characters. At about 900 characters per minute of natural speech, one minute starts near one credit.

Vapi usage, telephony and any fallback provider are separate. Use the TTS API pricing guide for Fish Audio credit planning.

How do you choose the fallback voice?

Choose a fallback that preserves the same role, language and rough pace as the primary Fish Audio voice. A technically successful fallback that switches from a calm support agent to a theatrical narrator is still a failed customer experience.

Keep the fallback voice in Vapi’s supported provider set and document it in the assistant version. Test the switch on a complete turn, including tool output and interruption, rather than only generating a sample sentence.

Alert when fallback usage rises. Frequent switching usually indicates provider configuration drift, exhausted credits or latency beyond the assistant budget.

If your shortlist includes Retell as another agent host, compare the Retell AI text-to-speech setup guide before choosing the orchestration layer.

What should the pilot cover?

Run at least three call types: a healthy short answer, a long tool-driven answer and a failed TTS response. Confirm the same voice, correct audio format, fallback path and interruption behavior in each.

Review transcripts and call recordings under your privacy policy. A provider integration can be technically healthy while the assistant remains poor at turn-taking.

Key takeaways

• Verify Vapi’s current custom TTS contract before building.
• Fish Audio supplies speech; Vapi owns orchestration, tools and fallback.
• Store voice identifiers with assistant versions and consent records.
• Measure p95 end-to-end latency and test failure fallback.
• Fish Audio and Vapi usage are billed separately.

Frequently asked questions

Does Vapi support custom TTS providers?

Vapi documents a custom TTS path. Verify the current provider contract and supported audio response format before connecting Fish Audio.

Is Fish Audio a native Vapi provider?

Do not assume native support. Check Vapi’s current provider list. If Fish Audio is not listed, use the documented custom TTS approach or a bridge you operate.

What should I test first?

Test one short turn through the real Vapi call path: caller audio, transcript, model answer, Fish Audio speech and playback. Measure the first audible syllable and interruption behavior.

How do I keep the custom voice consistent?

Store the Fish Audio voice identifier and script settings with the assistant configuration. Keep a consent record for cloned voices and version the voice used by each assistant.

How much does Fish Audio speech cost with Vapi?

Fish Audio consumes one credit for about 1,000 characters; API access starts with the $99 Studio pack. Vapi platform or telephony charges remain separate.