STT Accuracy
"STT Accuracy is how closely a speech-to-text engine's transcript matches what the caller said, usually scored as word error rate on real phone audio."
What is an STT Accuracy?
STT Accuracy is how closely a speech-to-text engine's transcript matches what the caller said, usually scored as word error rate on real phone audio. Related product: Voice Agents.
Reviewed by Jaffar Jawed, Co-Founder · Last reviewed
Deep Dive
Accuracy numbers on a vendor's website come from clean, studio-quality audio. The number that matters for a voice agent is how well the engine transcribes a caller on a mobile in a car, with an accent the model has not heard much, reading out a part number. This entry is about measuring that honestly and about what to do when the score is not good enough.
How it works. The standard score is word error rate (WER): the number of substituted, deleted and inserted words divided by the number of words in a human-corrected reference transcript. A transcript with 10 errors against a 100-word reference has a WER of 0.10. Because not all words matter equally, teams also track entity accuracy: whether names, phone numbers, addresses, dates and product names came through intact. Engines also emit a confidence score per word or segment, which the agent uses to decide when to read something back.
Why it matters for call outcomes. Every recognition error is a potential wrong action. A wrong digit in a callback number means the confirmation goes nowhere. "Cancel" misheard for "can't tell" closes a booking the caller wanted. Poor accuracy also inflates transfer rates, because the agent falls back to a person whenever confidence is low. Improving accuracy on the specific vocabulary of your business is often the single most effective tuning step after launch.
Example. A Shopify merchant selling bike parts runs a voice agent for order questions. The first week's transcripts show "derailleur" transcribed six different ways and order numbers wrong on roughly one call in eight. The team adds a custom vocabulary of 200 product terms, switches the agent to reading order numbers back digit by digit for confirmation, and enables noise suppression on the telephony leg. The following week's sample shows order numbers confirmed correctly on nearly every call and a lower transfer rate.
STT accuracy vs ASR. ASR is the component; STT accuracy is its measured quality. Two agents can use the same recognition engine and score differently because of microphone path, codec, noise handling and vocabulary configuration. Treat accuracy as a property of your deployment, not of the vendor.
What to measure.
| Measure | How to get it |
|---|---|
| WER on your own calls | Monthly sample, corrected by a person, re-scored after every change |
| Entity accuracy | Names, numbers, addresses and dates checked field by field |
| Accuracy by caller segment | Mobile vs landline, accent, language, time of day |
| Confidence calibration | Do low-confidence segments actually contain the errors? |
| Read-back correction rate | How often the caller corrects a confirmed value |
Botcadence voice agents target sub-second responses and pass the transcript to a person on handoff, so recognition quality shows up directly in what your team reads; see voice agents and the notes on telephony denoising, managed vs self-hosted.
Related Technology
See STT Accuracy in a Botcadence agent
Start on the Free plan and configure an agent that uses STT Accuracy on your own website chat or WhatsApp number. No card needed.
Get Started FreeWebsite chat live in about 5 minutes
Common Questions
STT Accuracy FAQs
Phone audio is band-limited and compressed by the carrier, and many calls arrive from mobiles in noisy places. Demos use wideband microphone audio in quiet rooms. Benchmark with recordings taken from your own line.
Pull 50-100 recent calls across times of day and caller types, have a person correct the transcripts, and store them as the reference. Re-run the engine against the same set after every configuration change.
Judge by outcome rather than a single number: if names, numbers and addresses are confirmed correctly at the read-back step and the low-confidence transfer rate is acceptable, the agent is usable. Keep improving vocabulary after launch.
Often yes with modern multilingual engines, but accuracy varies by accent. Measure WER per caller segment, add vocabulary that accented speakers use, and route to a person when confidence stays low across several turns.