The average no-show rate across US healthcare sits between 5-30 percent, and calling patients one by one doesn't scale. AWS just published a reference implementation that replaces the standard three-model voice pipeline with Amazon Nova 2 Sonic, a speech-to-speech model that processes audio natively without separate transcription, reasoning, and synthesis steps.
Why Chaining STT-LLM-TTS Fails for Healthcare Traditional voice agents chain a speech-to-text model, a text-based LLM, and a text-to-speech model. Each handoff introduces latency and drops context. The transcription step discards vocal cues like tone, hesitation, and urgency. The LLM never hears how the patient said something, only what they said. In healthcare, where a patient's anxiety or confusion should change how the agent responds, that loss matters. Nova 2 Sonic eliminates those handoffs. It processes speech natively in a single model, retaining vocal nuance across turns. The model picks up on hesitation or concern without losing context, handles background noise common in homes and clinics, and switches to a patient's preferred language mid-conversation without configuration changes. AWS claims multilingual support is built in.
Seven Tools and a System Prompt Drive the Conversation The agent uses Amazon Bedrock AgentCore with the Strands Agents SDK's BidiAgent class. The entire setup fits in a few lines of Python: ```python
from strands.experimental.bidi import BidiAgent
from strands.experimental.bidi.models import BidiNovaSonicModel model = BidiNovaSonicModel( region="us-east-1", model_id="amazon.nova-2-sonic-v1:0", provider_config={ "audio": { "input_sample_rate": 16000, "output_sample_rate": 16000, "voice": "tiffany", } }, tools=tools,
)
agent = BidiAgent(model=model, tools=tools, system_prompt=system_prompt)
await agent.run(inputs=, outputs= )
- **authenticate_patient** verifies identity via first name, last name, and last four SSN digits (three-attempt limit, then escalates)
- **confirm_appointment** and **cancel_appointment** update DynamoDB with idempotent checks
- **find_available_slots** queries up to three open slots via a Global Secondary Index
- **book_appointment_slot** uses DynamoDB conditional writes to prevent double-booking
- **record_health_update** captures four pre-visit items (conditions, allergies, accompaniment, concerns) one at a time
- **escalate_to_agent** generates a six-digit reference number and publishes an SNS notification for human callback The system prompt defines the conversation flow: authenticate, present appointment, handle confirmation/cancel/reschedule, collect health info, escalate on request. Nova 2 Sonic invokes tools as needed. You swap or extend tools without rewriting the agent.
## Serverless Deployment with CDK AWS provides an AWS CDK stack that provisions Cognito authentication, three DynamoDB tables, an SNS topic, and the containerized AgentCore runtime. A React frontend streams bidirectional audio over SigV4-signed WebSockets. Sample patient data is seeded automatically: John Smith (SSN last 4: 1234), Jane Doe (5678), Michael Johnson (9012). Deploy with `cdk deploy`, create a test user via Cognito CLI, then open the frontend at localhost:5173. The demo shows a complete rescheduling interaction: the agent verifies identity, confirms the appointment, finds morning slots, books Tuesday July 1st at 9:00 AM, and collects high blood pressure and penicillin allergy info. Spanish and Hindi demos are included in the repo. Clone the repository at `https://github.com/aws-samples/sample-Nova-Sonic-AgentCore-Healthcare-Call-Center.git`, customize the system prompt and tools for your domain, and test your own appointment flows.
---
Source: [Build a healthcare appointment agent with Amazon Nova 2 Sonic](https://aws.amazon.com/blogs/machine-learning/build-a-healthcare-appointment-agent-with-amazon-nova-2-sonic/)
Domain: aws.amazon.com
Comments load interactively on the live page.