A voice agent can transcribe a sentence about a broken furnace perfectly and then fail on the street it is standing outside.

Why are names harder than sentences?

Because recognition is not pure transcription; it is acoustics plus prediction. Ordinary language is highly redundant, so a model that half-hears a word can recover it from the words around it. A proper noun carries no redundancy: it can be almost anything, nothing predicts it, and a single misheard phoneme produces a different name rather than an obviously wrong sentence.

This is also why the errors are so much more damaging than their frequency suggests. A misheard ordinary word usually produces something visibly wrong that a human reading the transcript corrects instantly. A misheard surname produces a plausible name that nobody catches until someone calls it back and reaches a stranger.

Street names compound it by adding local knowledge, and the fix is local everywhere rather than local to one place. A Calgary line hearing "Sarcee" or "Deerfoot" needs those in scope; a Boston line needs "Worcester" and "Leominster"; a Houston line needs "Kuykendahl". A general-purpose model weights each of them against every other similar-sounding sequence in its training data, almost none of which came from the city the caller is standing in. The same applies to clinic names, product names, and the surname of the technician the caller is asking for.

Alphanumerics are the third category and the worst. Letter names in English were not designed to be distinguishable over a compressed 8kHz phone channel: B, C, D, E, G, P, T, V and Z are the same vowel with a brief consonant, and the consonant is what the codec throws away. Postal codes and policy numbers live squarely in that set.

What does a voice agent actually do about it?

It stops relying on getting the name right the first time. The reliable designs confirm rather than transcribe: read the value back, spell it where it matters, constrain the search space to a known list where one exists, and capture the audio so a human can resolve what the system could not.

There is a design temptation worth naming: making the agent ask everyone to spell everything. It works, it is accurate, and it is unbearable to talk to. The judgement is which fields genuinely have to be exact: usually the callback number and the address, rarely the name of someone already in the system.

The callback number is the field that deserves the most protection, because it is the one that makes every other error recoverable. A line that gets the name wrong and the number right can fix itself. The reverse cannot.

• Read it back. Repeating a name or address to the caller converts a silent error into a corrected one, and costs about three seconds.
• Constrain where you can. If the caller is an existing customer, the name is in a list of a few thousand rather than open. Matching against a known set is a completely different problem from open transcription.
• Spell the things that must be exact. Email addresses and postal codes are worth spelling. Names of returning customers usually are not.
• Keep the audio for the hard fields. A human resolving one ambiguous surname in four seconds beats a system guessing confidently.

Does a better model fix this?

It moves the number, not the shape. Recognition has improved enormously and continues to, but the open-set problem is structural: a system predicting from context cannot predict a word that context does not constrain. A design that assumes perfect transcription of proper nouns will fail on whatever the current error rate happens to be.

Which is the practical reason to be suspicious of accuracy claims quoted as a single percentage. Word error rate on read speech in a quiet room has almost nothing to say about surname accuracy on a compressed phone line with a van engine in the background. Both are real numbers and they measure different things.

The useful question for a vendor is not "how accurate is it" but "what does it do when it is not sure". A system that confirms, constrains and escalates is robust to its own error rate. A system that transcribes and hopes is not, at any accuracy.