Amazon Connect Custom Voice: Every Voice Option for an AI Voice Agent

People searching for an Amazon Connect custom voice usually want one of three things, and the setup for each is completely different.

1. A better built-in voice: a more natural or more expressive voice than the default, using what AWS already provides.
2. A third-party voice: a specific voice from a provider that Connect can call, such as ElevenLabs, with your own account and key.
3. A voice made for your brand: a custom voice built for your organisation rather than picked from a list.

This guide covers all three against the current Amazon Connect Customer documentation, which we checked on 10 November 2026. It focuses on voice agents on phone calls. If you are still building the agent itself, start with How to Add an AI Voice Agent to Amazon Connect, where voice is Step 8, and come back here when you want to choose or change the voice.

Prices change, so we do not quote any here. Check the Amazon Connect and Amazon Polly pricing pages and your provider's plan before you commit.

For most new voice agents this is the right place to start, because it is the default and needs no extra accounts. Two settings work together.

1. On the bot, open the speech configuration, set the voice provider to Amazon Connect agentic voice and the speech model preference to Advanced. AWS describes this as a streaming recognizer for lower latency and natural turn-taking.
2. In the flow, add or open a Set voice block, keep Amazon Connect agentic voice as the provider, choose the language (this filters the voice list), pick a voice, and use the listen-to-sample link before you commit.
3. Save and publish the flow.

If callers switch languages, AWS offers multilingual voices (Katie, Blake, Brooke, Ronald and Gemma at the time of writing) that can follow the caller from a single bot locale. For a voice in a language those do not cover, build a separate bot for that locale and use a voice made for it, setting the language explicitly. Details are in the agentic voice configuration guide.

The Set voice block also exposes Amazon Polly voices and engines. This is the long-standing route and still useful when you want a specific named voice or a standard prompt voice for non-agent flows.

• The block's default is Joanna in a conversational style. Override speaking style lets you choose neural or generative engines for voices that support them.
• Neural voices improve pitch, inflection and tempo. Generative voices are described by AWS as the most human-like and adaptive Polly voices, and they are charged.
• Not every voice supports every engine. A voice such as Ruth does not support the standard engine, so a flow that does not specify an engine fails over to the error branch. If a later prompt plays after an error branch, it falls back to Joanna on the standard engine.
• You can set language, voice, engine and style dynamically, but only in consistent combinations: if the language is dynamic the voice must be too, and if the voice is dynamic and the style is overridden, the engine and style must be dynamic as well.
• Instances created before October 2018 that moved to a service-linked role may need an extra permission, polly:SynthesizeSpeech, on the service role to use the generative engines.

With Amazon Lex V2 bots, the language in the Set voice block must match the bot's language model, and for a non-US English voice you must choose Set language attribute. Otherwise the Get customer input block errors. Reference: Set voice.

Nova Sonic is configured as the bot's speech model rather than as a voice in the Set voice block alone. Set the bot locale's model type to Speech-to-Speech with Amazon Nova Sonic, build the language, and then in the flow choose Override speaking style, select Generative, set the voice provider to Amazon and pick a compatible voice.

AWS lists the launch voices as Matthew (en-US), Amy (en-GB), Olivia (en-AU) and Lupe (es-US). That is a short list, so check it against your language and brand before you design around speech-to-speech. Reference: Configure Amazon Nova Sonic speech-to-speech.

Connect can call an external text-to-speech provider from the Set voice block. This is the route for a voice you have chosen or created in the provider's own tools. The setup has more parts than the others, so do it in this order.

1. Store the provider's API key in AWS Secrets Manager, in the same Region as your Connect instance.
2. Use a customer-managed KMS key for that secret. AWS says you cannot use the default aws/secretsmanager key.
3. Attach a resource policy to the secret that lets the Connect service principal call GetSecretValue, limited to your instance and your account.
4. Attach a resource policy to the KMS key that lets Connect decrypt, limited to your instance, your account and that specific secret.
5. In the Set voice block choose the third-party provider, then set the model, voice, Secrets Manager ARN and language manually. The language must be one the provider's voice supports.
6. Save, publish and test.

{
  "apiToken": "<your provider API key>",
  "apiTokenRegion": "eu"
}

By default Connect calls the provider's standard endpoint. The apiTokenRegion field lets you point to another provider Region, and AWS lists us, eu and in options for ElevenLabs. Aligning the provider Region with your instance is also the check AWS suggests when a third-party voice sounds slow.

A note on cloned or custom voices from a provider: the Connect documentation only asks for a provider voice value. Whether a voice you created or cloned in your provider account can be used, and on what terms, is a question for that provider's documentation and licence. Get written consent from anyone whose voice is cloned. Reference: Configure third-party text-to-speech providers and Managing secrets and resource policies.

Recognition is separate from voice. If callers are poorly understood in a language or accent, you can change the speech-to-text provider on the bot instead. AWS documents Deepgram as a third-party option, set on the bot locale: choose the Speech model, set the model type to Speech-to-Text, pick the provider, enter the model ID (case-sensitive, so copy it exactly from the provider) and the Secrets Manager ARN, then choose Build language so the change becomes active.

The same secrets and KMS rules apply. AWS says no flow or Lambda changes are needed, errors appear in the logs, and a slow response should prompt you to check the provider Region. Reference: Configure third-party speech-to-text providers.

Amazon Polly Brand Voice is a custom neural voice created for a single organisation. The Connect documentation confirms that Brand Voices associated with your account are charged, which tells you they are a supported idea in Connect. What we could not find is a documented step-by-step for choosing one in the Set voice block, and it is built through an engagement with AWS, not a self-service setting.

If a distinctive brand voice is a firm requirement, ask AWS two questions before you plan the project: how the voice is selected in a Connect flow, and what it costs per character. For many teams a well-chosen built-in voice or a third-party voice gets there faster and is easier to change later.

A good voice still reads badly written text badly. For agentic voice, AWS's guidance is to pass natural, well-punctuated sentences and avoid markdown, emoji and brackets, because the engine reads symbols aloud. A few specifics worth knowing:

• Normal capitalisation matters. The engine treats ALL CAPS the same as a spell tag, so ACME BANK is read letter by letter. Write Acme Bank if you want it said as a name.
• Use the spell tag for codes and reference numbers, for example <spell>TKT4829XB</spell>, and use break tags only when natural punctuation does not give the pause you want.
• Speed, volume, emotion and laughter tags exist. Emotion tags are described as a beta capability. Start without tags and add one only where a test call shows a problem.
• Unclosed or malformed tags are read aloud as text, so validate them before they reach the caller.

For prompts played with Polly voices in non-agent flows, Connect supports SSML for pronunciation, volume and speed. Reference: Agentic voice best practices.

Whichever you choose, decide by listening. Call the number from a mobile phone in a noisy place and read through your real prompts: names, phone numbers, dates, reference codes and a long answer. Compare two voices on the same script. Also test the failure case, because if the voice or provider fails the call should reach a person, not silence.

For the architecture behind the voice, including latency, tools and escalation, read our architecture and build guide, and for how the agent connects to Connect, see AI voice agent integration with Amazon Connect. To have a voice agent built and tuned for you, see our AI voice agent development service.