Speech recognition accuracy
in noisy environments.

The first question every CTO asks about speech analytics is "So how accurate is it?" The honest answer is more complicated than a single number — and a vendor who gives you one number with no caveats should already raise a flag.

How accuracy is actually measured

The base metric is WER (word error rate) — the share of words transcribed incorrectly. Word-level accuracy = 100% − WER. The key point: WER is always measured on a specific type of audio. The same model will post different numbers on a studio recording, a phone line, and a wearable badge on a noisy sales floor.

Realistic numbers for real-world audio

If someone promises "99% in any conditions", it is either a lab benchmark on clean speech or plain marketing. On a floor with music and parallel conversations you cannot cheat physics: even a human transcriber loses some of the words.

What actually drives accuracy

• Distance to the microphone. A badge on an employee's chest hears them almost perfectly and the customer less well — so the expert's lines transcribe more accurately than the customer's.
• Overlapping speech. When people talk over each other, error rates climb for any model. Good systems flag those segments instead of passing them off as clean text.
• Domain vocabulary. SKU names, promo mechanics and trade slang have to be learned from your data — otherwise "trade-in" comes out as "training".
• Codec and channel. Narrowband telephony cuts off frequencies; modern codecs (Opus) transcribe noticeably better.

Why WER is not the metric that matters most

For the business, what matters is scoring accuracy, not transcript accuracy. A checklist of "stated the price", "offered the promo", "closed politely" is robust to isolated recognition errors: you do not need every word transcribed perfectly to score a line item correctly.

A vendor question checklist

• What is the accuracy on our recordings, not on your demo rig? (Ask them to run a pilot pool.)
• How are low-audio-quality segments flagged? Do they feed into scoring?
• How does the model learn our vocabulary: SKUs, promos, competitor names?
• What score agreement do you reach with our auditors after calibration? How do appeals work?
• Does the model keep learning from our corrections?

How we handle this ourselves — in the Frontline FAQ and the QA methodology.