What is speech-to-text?

Speech recognition works on probability, choosing the most likely sequence of words given the sound and given what usually follows what.

That is why it handles ordinary conversation well and unusual surnames badly. Nothing in the language makes a rare name more likely than a common word that sounds like it.

Phone audio makes it harder than a podcast microphone would. Narrow bandwidth, compression and background noise all cost accuracy.

Speech recognition is usually scored by word error rate. You count the words the system got wrong, added or dropped, and divide by the number of words actually spoken. Say a caller speaks 100 words and the transcript has six mistakes. That's a 6 percent error rate, which sounds excellent.

Now look at which six. Suppose they're "um", "the", "gonna" and three other fillers. Then the transcript is as good as perfect. If two of them are digits in the callback number and one is the caller's last name, you can't return the call and you can't look her up. Both transcripts score 6 percent. Only one of them is any use to you, and that's the real weakness of a single accuracy figure: it treats every word as equally valuable when a phone call has maybe ten words that matter.

Start with the audio. Phone calls are sampled at 8,000 times a second, far below what a podcast or a video meeting uses, and a recognizer built for phone audio tends to behave differently on it than a general one. Ask the vendor which kind is in use on your calls.

Next, ask whether you can supply your own vocabulary. Your street names, your staff's names and your product names are exactly the words a general model has never seen. Some engines let you list them so they're favored when the sound is close.

Finally, ask how you'd audit it. You want to be able to put a transcript next to what was actually said for a handful of calls, at least during setup. If a vendor can't show you a way to compare the two, you're being asked to take the accuracy on trust.

GreetKeeper publishes no accuracy figure, because a single number would describe one audio condition and mislead about every other.

Where recognition errors are expensive, the fix is confirmation rather than optimism. The assistant can read a phone number back in groups and spell a surname before ending the call.

Both sides of the call are transcribed, so when something goes wrong you can see whether the caller was misheard or the assistant simply answered badly.