Part numbers are where voice AI breaks

A foreman says "two-three-five, sixty, three-five-three-hundred". Your quotation needs 235-60-35300 and nothing else. One wrong digit is a different part, a wrong shipment, a freight bill for something nobody ordered and a customer who does not call again.

General voice assistants are built for sentences, where a wrong word is repaired by the words around it. A part number has no words around it. Every character carries the whole meaning, the digits are spoken in groups the caller chooses, and letters like B, D and E sound alike on a phone line with a generator running behind the caller.

85% is not good enough for a part number.

Even a speech model built for the job gets about 85% of mixed letter-and-digit sequences right; that is Speechmatics' own published figure for its model. For sentences that is excellent. For a thirty-line enquiry read out on the phone it means four or five wrong lines, and nobody knows which.

Why part numbers are hard for speech recognition

• There is no grammar to fall back on: 6D140 is not a word, and nothing predicts the next character.
• Callers group digits their own way: "twenty-three, five sixty" and "two three five, six zero" are the same number.
• Brands write the same part differently: 328-0730 and 3280730 are one CAT part.
• Phone audio is narrow: 8 kHz, with the high frequencies that separate F from S cut off.
• Callers switch language in the middle of a number, and count in the one they learned first.

Tell the recogniser what to expect

Before the call starts, the desk hands the recogniser up to fifty terms for this call: your brands, product families, the customer's name and the codes it is likely to hear. A term heard the same way on two calls (or once, for a code) joins that list automatically, and anyone on your team can remove it again. Over weeks, the desk learns your catalogue's vocabulary from your own calls.

Listen twice to what was spelled

The desk keeps the last minute of audio. When the caller spells something or says a number, that stretch is transcribed a second time, in isolation, before the agent reads it back. The agent never repeats a part number it has only heard once in passing, in the middle of a sentence about something else.

Read it back, digit by digit

The agent reads the number back in the caller's own grouping and waits for a yes. Then it looks the part up in your catalogue: what it is, whether you can supply it, and its equivalents from other brands. It can tell the caller that you carry the filter and that an aftermarket equivalent exists. Prices, costs and suppliers never reach the phone: any lookup result that mentions them is withheld from the agent entirely.

Let the recording correct the record

When the call ends, the recording is listened to again, whole. A misheard e-mail address or code on the call's record is corrected from it. A Cyrillic К or А typed into a part number on a Russian keyboard is turned into its Latin twin, so the catalogue finds it.

A checklist for any voice AI you evaluate

1. Ask it to take a thirty-line enquiry by phone and compare the transcript with the list.
2. Spell a number in two groupings and see whether it reads back the same part.
3. Say a number in Hindi or Russian and see what lands on the record.
4. Ask it for a price and see whether it gives one.
5. Check whether the record is corrected after the call, or only written once.

What this means for a parts trader

• Night calls become orders instead of voicemails nobody can read.
• The number on the CRM card is the number the customer said.
• Your team checks the enquiry in the morning instead of calling back to ask again.

The systems behind this are TERRA-109 READ and TERRA-110 LOOKUP, working inside TERRA-101 CALL, with TERRA-306 CATALOGUE as the list every number is checked against.

Questions