Automatic Speech Recognition

Post Call Analytics
Automatic Speech Recognition

Inspiration

Contact centers are sitting on "dark data"—millions of voice hours that just vanish. We realized that companies were only auditing about 1% of their calls manually. It felt like a massive waste of insight. We wanted to close that gap and actually hear what customers were saying, at scale, without the manual bottleneck.

What it does

We’ve automated the analysis of every single call. The system transcribes audio, tags the reason for the call, flags sales leads, and catches complaints instantly. We can now process a 17-minute call in about 25 seconds. It turned a blind spot into a dashboard that sub-managers get via email every morning, all for about 0.8 cents per call transcribed and a total of 8 cents per call analyzed.

How we built it

We went with a cloud-native AWS stack built for speed:

The Engine: Amazon ECS clusters running on GPU-equipped Spot instances for high-speed ASR.

The Brain: AWS Bedrock (Prompt Management and Flows) to handle multi-label classification and Amazon Nova for real-time context detection.

Orchestration: S3 triggers a workflow through Lambda and SQS. We store everything in RDS as structured Q&A pairs for easy querying.

Challenges we ran into

Scaling is where things get messy. When you're processing this much volume, "edge cases" become daily occurrences. We had to build a very resilient architecture using SQS and Dead Letter Queues (DLQs) to ensure no call was lost or double-billed. Balancing deep AI analysis with a strict budget was a constant tug-of-war.

Another challenge was to debug the prompts of the multi label classifier and the analysis flows, this allows the complete process to function efficiently.

What we learned

We learned that you don't need to go "all-in" on every call to be effective. We implemented probabilistic execution: 100% depth for high-stakes residential services, but smart sampling for executive audits. It kept the budget sustainable. We also learned that in a massive pipeline, you have to design for failure from day one—if it can break, it will. We also have to learned how to use correctly the prompts and flows.