Voice AI technology

Getting Names, Numbers and Addresses Right by Voice on Calls

Why a voice agent mishears digits, surnames and street names, and the read-back, chunking and fallback techniques that make capturing these details dependable.

Call editorial team · Published · 6 min read

A voice agent can have an excellent conversation and still fail on the one thing the business needs: a phone number written down wrongly, a surname misspelled or a street that does not exist. These details are short, have no context to rely on and are unforgiving, because a single wrong digit makes the whole item useless.

A person writing down details while on a phone call
Illustrative image

This guide explains why these details are hard to capture by voice and gives a set of techniques that work. For the audio side of the problem, read Why Phone Audio Is Hard for Speech Recognition.

Why names, numbers and addresses are hard

Four sources of error

  • Similar soundsDigits such as five and nine, or letters such as B and D, are easily confused on a phone line.
  • No contextA word in a sentence is guessed from the words around it. A reference number has no such help.
  • Open vocabularySurnames, streets and company names are endless and often not in a dictionary.
  • Speaking habitsPeople group digits differently, say double or triple, or pause in odd places.

Numbers also have several spoken forms. A caller might say one oh five, a hundred and five or one zero five. A date can be given in different orders. A recogniser must map each to a standard form, and the mapping may differ from what the caller meant.

Capturing numbers

Ask for numbers in chunks

Instead of asking for the whole number at once, tell the caller you will take it in parts. This shortens each turn, reduces errors and makes repair simple. Pausing in expected places helps the recogniser, too.

A reliable number capture

  1. Ask for the first groupFor a long reference, ask for the first few characters.
  2. Read back that groupRepeat it back and wait for a yes or a correction.
  3. Continue with the next groupRepeat the cycle for the rest.
  4. Read the whole thing backConfirm the complete number once at the end.
  5. Offer a second routeIf it fails twice, take it by message or send a form.

Be patient with silence

A caller reading digits from a card will pause between groups. If the agent treats each pause as the end of the turn, the number is cut in half. A longer silence threshold for these questions prevents it; see Voice Activity Detection.

Validate against known rules

Many numbers have a known length or pattern. If the agent hears seven digits where eight are expected, it can ask the caller to repeat rather than accept it. Where a number can be checked against your own records, a quick match confirms that the caller said what you think.

Capturing names

Names are harder because they are not just words; they are identities. A person is quick to notice when their own name is wrong. Treat names with care.

A frequent compromise is to take the name as heard and mark it for a person to verify before any official document is created. That moves the effort to where it can be done reliably.

Capturing addresses

An address has parts: number, street, area, postcode or equivalent. Capture them one at a time. Postcodes and zip codes are particularly useful because a valid code narrows the street to a small set, and a validation list can catch impossible combinations.

Where possible, check the address against a service that you trust and let the agent repeat the matched version. If no such lookup is available, read it back and let the caller correct a part. Never assume that the version heard is the real one.

Voice versus a second channel

Sometimes the right answer is not to capture the detail by voice at all.

Capturing a detail by voice or another way

By voiceBy another route
SpeedQuick when it worksSlower; the caller moves to a form or a message
AccuracyVariable on a phone lineTyped details are usually clearer
Caller effortLow, but repeated errors annoyModerate; some will not complete it
Best forShort items with a read-backLong strings, rare names, email addresses

An email address is a good example. It combines letters, symbols and unusual words in one string, which is among the worst items for voice. A booking page or a form is more dependable. Call does not send text messages by itself, so check what channels your business has set up before you promise that a link will arrive; a spoken address for a web page or a message to your team may be the practical route.

Wording that helps

The way a question is phrased changes the answer. Compare asking for your phone number, which invites a rapid stream, with asking for the first three digits of your mobile number, which invites a short, careful answer. The second is slower on paper but faster in practice, because it avoids the repeat.

Read-backs deserve equal care. Repeat the digits in the same groups the caller used, with the same rhythm, and end with a clear question such as is that right? Avoid reading the whole string in one breath. Where the caller corrects a single digit, fix that digit and read the group again, rather than restarting the whole number.

Finally, set expectations early. A short sentence such as I will read each part back to check it tells the caller why you are slowing down, and they will cooperate rather than rush. A caller who understands the process is more patient than one who is surprised by it, and more likely to give a correction at the right moment.

Setting this up in Call

In Call, write the questions for the agent's script in short, single-purpose steps and give it the read-back wording as exact text where it matters. Use the transcriber settings to give callers time to pause. Keep in mind that the agent cannot book or confirm appointments on the call; it takes a request, and a team member confirms. Capture details so that the person who follows up can verify them, and use the help desk to record, assign and resolve those requests. Afterwards, search the call transcripts for the details that were repeated, to see where callers struggled. More on appointment requests is in Taking Appointment Requests with an AI Receptionist.

Before launch, test these cases

  • A long reference number read with pauses
  • A phone number said in different groupings
  • A surname with an unusual spelling
  • A street name that sounds like a common word
  • A caller who corrects a digit after the read-back
  • A caller who gives up and asks for a person
  • A detail the agent cannot capture after two tries

Frequently asked questions

Why does the agent mishear digits?

Digits often sound alike on a narrow phone line, and a number has no surrounding words to help. Chunking, read-back and longer pauses improve accuracy.

Should callers spell their names?

Only when needed. Spelling by voice is error-prone, so offer it as a fallback, and mark names for a person to verify before they are used in official records.

Is it better to collect an email address by voice?

Usually not. Email addresses mix letters and symbols and are hard to hear correctly. Use a web form or ask the team to follow up where accuracy matters.

How do I stop the agent cutting off a caller reading a number?

Increase the silence threshold and ask for the number in small groups, so natural pauses are not mistaken for the end of the turn.

Can the agent check an address?

Only if it is connected to a lookup you trust. Otherwise, have it read the address back for the caller to confirm, and let a person verify before use.

What should happen when capture fails?

After two failed attempts, change method: offer a message, a form or a person. Repeating the same question a third time rarely works.

Capture caller details more reliably with Call

In the demo console you can try incoming calls in a simulation.

Related guides