Amazon Connect Text to Speech: Custom Voice API (2026)
Compare Amazon Connect built-in TTS with custom voice architectures using Fish Audio, including connectors, PCM playback, latency and compliance.
Amazon Connect text to speech can use its native voice options or a current third-party TTS provider path. Fish Audio is a separate speech-generation API: choose it only when the required provider fields, authentication and audio contract can be satisfied without bypassing Amazon Connect's documented configuration.
AWS currently documents a Set voice block with a third-party provider, model, voice, Secrets Manager ARN and language. That is a real configuration surface, but it is not evidence that Fish Audio appears in the provider list. Confirm the exact supported providers and model identifiers before calling this a native integration.
What does Amazon Connect handle versus Fish Audio?
Amazon Connect owns contact flows, routing, agents, queues, call control and telephony. Fish Audio generates expressive speech from text. They become one system only through a documented provider contract or an external media component that Amazon Connect is allowed to call.
Keep the responsibilities explicit:
- Amazon Connect: contact flow, number, routing, transfer, queue and agent experience.
- Current TTS provider configuration: model, voice, language and credential reference.
- Fish Audio: voice identity, speech generation, pronunciation controls and character usage.
- AWS security layer: Secrets Manager, KMS, IAM and CloudWatch controls.
Read the official Amazon Connect third-party TTS guide before implementing. The configuration sequence and supported fields are the authority.
How does the third-party TTS configuration work?
AWS documents five steps: open the contact flow, add or select a Set voice block, choose a third-party TTS provider, specify the model, voice, Secrets Manager ARN and language, then save and publish the flow.
That sequence creates three pre-launch questions:
| Question | Why it matters | Evidence required |
|---|---|---|
| Is Fish Audio an accepted provider/model? | An API being compatible in theory does not make it listed in the console | Current AWS provider and model documentation |
| Does the voice ID map to the provider field? | A browser voice identifier may not be valid in Amazon Connect | A successful staging flow using the exact ID |
| Can AWS retrieve the secret? | The runtime needs the key without exposing it in the flow | Secrets Manager policy and KMS decrypt test |
If Fish Audio is not an accepted provider, do not force an undocumented field value. Use a supported custom-media architecture only if AWS documentation and your security review allow it.
How should credentials and language be handled?
Store the provider API key in AWS Secrets Manager and let Amazon Connect retrieve it under a least-privilege resource policy. The KMS key must allow the required decrypt operation. Never paste the key into a contact-flow attribute, Lambda source file or public runtime setting.
The language field should match the generated speech, not merely the caller's menu language. A multilingual voice still needs pronunciation tests for names, product terms and numbers. Keep one approved voice and language mapping per queue or campaign.
Log the secret ARN and provider request identifier, never the secret value. Rotate credentials through Secrets Manager and retest the contact flow after rotation.
What audio and latency constraints matter?
The call path ultimately reaches telephone audio. Confirm the sample rate, encoding and transport the current provider integration returns to Amazon Connect. Do not assume a studio WAV or 24 kHz browser file is accepted unchanged.
Measure the complete response time from the caller's last syllable to the first audible word. Include:
- speech recognition finalization;
- contact-flow logic and any Lambda call;
- LLM or knowledge retrieval;
- Fish Audio time to first audio;
- network transit, buffering and playback.
Record p50 and p95. The telephony text to speech API guide covers PCM, SIP and sample-rate failures in more detail.
How should barge-in and fallback work?
The contact flow must stop audio when the caller speaks or when the agent takes over. The provider request should be cancelled where the current contract supports it, and stale audio must not play after the next turn.
Amazon Connect's native fallback and error path should remain available. If the third-party provider is unconfigured, times out or returns malformed audio, route to:
- a short pre-recorded clarification;
- the native Amazon Connect voice path where permitted;
- a controlled human transfer.
Test each route before production. See voice-agent TTS fallback for fault classes and transcript-aware recovery.
How much does the stack cost?
Fish Audio API access comes with the $99 Studio pack and 15,000 credits. One credit is approximately 1,000 generated characters. Amazon Connect, telephony, Lambda, Secrets Manager and other AWS services are billed separately.
Plan speech volume by generated characters plus retries. A one-minute reply of about 900 characters commonly uses one credit, but an interrupted or regenerated answer can cost more than the final transcript suggests.
Use TTS API pricing for character calculations. Do not treat an Amazon Connect telephony rate as an all-inclusive speech price.
What should the launch test include?
Run a staging contact flow with a real Amazon Connect number and the same region and IAM policy as production.
Test:
- short greeting and long tool-driven answer;
- caller interruption at the beginning, middle and end;
- mobile, landline and poor network audio;
- names, dates, account identifiers and two languages;
- secret rotation and provider timeout;
- native fallback and human transfer.
Keep the successful voice ID, model, language and secret reference in release notes. After launch, monitor first-audio p95, generation failures, fallback rate, abandoned calls and generated-character cost per completed contact.
Key takeaways
- Amazon Connect documents a third-party TTS provider configuration; Fish Audio must meet that exact contract.
- No native Fish Audio add-on should be claimed without current official evidence.
- Store credentials in Secrets Manager and verify KMS permissions.
- Test time to first audio, interruption and fallback through the real contact flow.
- Budget Fish Audio credits and AWS telephony separately.
Related telephony guides
- Telephony text to speech API - codecs, PCM, latency and platform boundaries.
- Twilio text to speech - a comparable media-stream architecture.
- Voice-agent TTS fallback - provider-neutral failover.
- Telnyx text to speech - REST, WebSocket and live-call playback boundaries.
- Plivo text to speech - Speak, Play and streaming routes.
- Vonage text to speech - NCCO talk and custom streams.
- Asterisk text to speech - self-hosted files and AudioSocket.
Frequently asked questions
Can Amazon Connect use Fish Audio for text to speech?
Amazon Connect supports a current third-party TTS configuration where its provider contract allows it. Fish Audio must match the supported provider, model, voice and credential fields before you plan a native configuration.
Does Fish Audio provide an Amazon Connect add-on?
No native Fish Audio add-on is claimed here. Use only a documented Amazon Connect third-party provider path or an approved external media architecture.
What belongs in the Amazon Connect configuration?
The AWS guide uses a Set voice block, third-party provider, model, voice, Secrets Manager ARN and language. Keep the credential reference in AWS and verify all values in staging.
How should phone audio be tested?
Measure from the caller's last syllable to the first audible reply. Test the real Amazon Connect phone path, interruptions, names, numbers and fallback on mobile and landline audio.
How is Fish Audio billed with Amazon Connect?
Fish Audio consumes prepaid character credits. Amazon Connect, telephony, Lambda and other AWS services are separate charges.