Normalized claim
Time to first audio: 1.4 seconds
Amazon Nova 2 Sonic achieved Time to First Audio of 1.39 seconds.
Loka built a low-latency speech-to-speech voice agent for automotive dealerships using Amazon Nova 2 Sonic (via Amazon Bedrock)
Use case typeVoice automationUpdated Jun 24, 2026
Loka built a conversational AI voice agent for automotive dealership calls using Amazon Nova 2 Sonic on Amazon Bedrock to reduce latency and improve natural turn-taking. The solution uses a serverless, event-driven architecture with LiveKit Agents on AWS, Amazon ECS, Amazon RDS, IAM governance for prompt changes, and Python tools for inventory search, appointment booking, and customer lookup.
- Organization
- Loka
- Industry
- Automotive
- Location
- United States
- Published
- June 2026
Reported outcomes
Time to first audio: 1.39 seconds
Time & speed
Speech reasoning score: 87 scoreOverall judge score: 2.4–2.7 scoreOverall judge score after prompt refinement: 2.4–2.7 scoreOverall judge score after second prompt iteration: 3.8 score
Why do we believe this?Outcome claims, sources, and evidence checks
Normalized claim
Speech reasoning score: 87 score
Amazon Nova 2 Sonic achieved a speech reasoning score of 87.0.
Normalized claim
Overall judge score: 2.4-2.7 score increase
The overall score increased from 2.4 to 2.7.
Normalized claim
Overall judge score after prompt refinement: 2.4-2.7 score increase
The overall score increased from 2.4 to 2.7.
Normalized claim
Overall judge score after second prompt iteration: 3.8 score
The second iteration achieved 3.8 out of 5.0.
Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks
- Customer
- Loka
- Provider
- AWS
- Maturity
- Production
- Linked source
- AWS Machine Learning Blog
The article says dealerships are already using the technology in production
Customer identity supportedSource describes one deploymentMaturity supported
Primary read
Use case focus
Showing 2 of 2
- 1Voice automation
- 2Customer service agent
- Loka used Amazon Nova 2 Sonic for end-to-end speech-to-speech processing.
- The team built a serverless, event-driven system using LiveKit, containerized agents on Amazon ECS, Amazon RDS persistence, IAM controls for prompt governance, and tool-based integrations with dealership systems.
- Prompts were treated like code and iteratively optimized with an automated LLM-as-a-judge evaluation pipeline.
- The model achieved 1.39 seconds time to first audio and 87.0 on Big Bench Audio.
- Overall judge score improved from 2.4 to 2.7 and prompt refinements raised it to 3.8.
- The article says dealerships are already using the technology in production.
Architecture
A serverless, event-driven voice-agent architecture used LiveKit for web and SIP connectivity, Amazon ECS for containerized LiveKit Agents, Amazon RDS for persistent dealership and conversation data, Amazon Bedrock for Amazon Nova 2 Sonic access, AWS IAM for prompt-governance controls, and OpenTelemetry for observability.
Sources & evidence1
Evidence: Medium50/100Evidence strength
- Customer explicitly identified
- Deployment status explicitly supported
- Quantified outcome available
- Technical implementation details available
Type: Blog PostPublished: Jun 24, 2026Publisher: AWSEvidence: VendorConfidence: Medium
AI-generated summary. Verify important details with the linked sources before relying on this case.
Explore related AI use cases
Customer evidence
Provider evidence
Browse the catalog
Was this useful?
Community
Comments
No published comments yet.