A real-time speech-to-speech AI service on AWS

A GovTech business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas, each mapping to one link in the voice chain, plus the underlying need for the whole thing to be reproducible.

Cloud Combinator delivered the service as four clear milestones. Each stage was wired up and validated on its own before being chained into the end-to-end flow, and every piece of infrastructure was defined in AWS CloudFormation so the whole environment could be stood up, torn down, and rebuilt consistently in the client's account.

The Proof of Concept met its success criteria across all three focus areas: audio in, accurate transcript, a context-aware answer, and natural speech back out, delivered as a single serverless pipeline that the client now owns in its own AWS account.

• Accurate transcription — Authenticated callers send audio and receive a reliable transcript from Nova Sonic speech-to-text, setting a dependable foundation for everything downstream.
• Relevant answers — An Amazon Bedrock large language model turns each transcript into a coherent, domain-relevant response, with a prompt template and response schema in place ready for further tuning.
• Natural voice output — Nova Sonic text-to-speech returns human-like audio rather than robotic playback, so the experience matches the product ambition.
• Repeatable environment — Because the whole stack is defined in CloudFormation, the service can be rebuilt consistently and handed over cleanly to the client's account.
• Handover — A runbook, handover deck, and recorded demo left the client's team able to operate and extend the service without dependence on Cloud Combinator.

The Proof of Concept is the start of the client's AI capability, not the finish. With a proven serverless pattern and a clean handover in place, the natural next steps are hardening the pipeline towards production, fine-tuning the language model on the client's own material, and using caller metadata to tailor responses further. The client now has a working, ownable foundation to take that product forward on AWS.