The Problem Nobody Wants to Admit
You have thousands of call recordings sitting in some server. Maybe Zendesk, maybe a custom system. They're data, technically, but they're not actionable data—which means they're basically noise. I spent three years building dashboards in Looker Studio and Power BI for companies that had perfectly structured databases but couldn't answer the simplest question: Why are our customers actually upset?
So I decided to build something. Not because the market needs another sentiment analyzer—it doesn't—but because I needed to prove to myself that raw audio could become the kind of clean, queryable insight that makes a CFO actually change something. I'm not sure this solves the problem at scale, but the prototype works and that's where everything starts.
How This Actually Works
- Whisper transcribes your audio into text—OpenAI's model, trained on 680,000 hours of multilingual audio, which frankly feels like overkill until you try it on Colombian Spanish with background noise
- You get a transcript that's actually readable. Still not perfect.
- BERTopic clusters those transcripts into themes without you needing to label anything manually, which is the part that feels like magic but is really just statistical pattern matching
- Sentiment scoring happens alongside topic extraction—so you don't just know customers complained about billing, you know they were furious about it
The workflow is straightforward, but I'll be honest: the infrastructure requires some setup. You need Python, Streamlit for the interface, and enough compute to run this without waiting forty minutes per call. I'm not convinced everyone actually needs this level of sophistication—a lot of companies would benefit more from just listening to ten calls a week—but that's not what I built.
What This Actually Tells You
Here's where it gets interesting and also where I'm genuinely uncertain about the value. When you process fifty calls and see that 34% involve complaints about wait time, 28% about product confusion, and 12% about pricing, what do you do with that? In Looker Studio, I'd build a nice visualization. In Power BI, I'd set up an automated refresh. But this data wants something different—it wants narrative context.
A sentiment trend that drops 15 points week-over-week looks alarming in a chart. When you actually listen to three of those calls and hear the same frustrated customer service rep using identical dismissive language, suddenly you're not looking at a metric. You're looking at a training problem. The data led you there, but the insight came from human judgment. This tool accelerates that journey, but it doesn't replace it.
I've built fifty dashboards that stakeholders ignored because the data was too clean, too abstract. This approach feels different because it's grounded in actual human language. Whether that translates to better decision-making depends on your organization's ability to act on nuance.
The Part I'm Still Figuring Out
Deploying this is harder than building it. Whisper costs money at scale—OpenAI's API charges per minute of audio—and BERTopic requires decent hardware to run efficiently on large datasets. For a mid-size call center processing 500 calls daily, you're looking at infrastructure costs that might compete with hiring a dedicated QA person who actually listens to calls. I'm genuinely not sure which is the better investment.
Also, there's the question of what happens when your sentiment analyzer catches something serious—a customer threatening to leave, or worse, a rep crossing ethical lines. Do you build automated alerts? Manual review queues? Who actually acts on this information? The technical problem is solved. The organizational problem is
