Case studies · VIBE mode
A 20-second voice memo from a phone counter, turned into a cited stock record
A store's demand sheet, filled from a counter memo with no typing: eight fields, each pointing at the second it was said.
The problem
A phone store keeps a demand sheet: who wanted what, in which variant, when they will come back. Staff fill it in by hand after the customer leaves, if they remember. The memo below is the kind they record instead.
"A customer wanted the 256 GB phone in blue. We only had the 128. She will come back on Saturday. Call her as soon as the blue one arrives. Two more people asked for the same model today. And we are out of fast chargers."
What VIBE did
The store showed Vlocalhost its sheet once, as a header row. A person checked the eight fields and approved them as a contract. The memo then went to the app's local API, was transcribed on the laptop, and came back as a record:
| Field | Value | Said at |
|---|---|---|
| Product | 256GB phone | 00:00:02 |
| Variant wanted | 256GB in blue | 00:00:02 |
| Customer returns | 10-10-2026 (“on Saturday”) | 00:00:09 |
| Call customer | Y | 00:00:09 |
| Also asked | 2 (“Two more people”) | 00:00:14 |
| Out of stock | fast chargers | 00:00:14 |
Store code and date came from the contract and the recording, not from the model.
What we fixed on the way
The first run dropped two correct values. The model answered “next Saturday” where the speaker said “on Saturday”, and “2” where they said “Two”; our check demanded the exact words. We changed it to check meaning, and to convert only the words actually spoken, because “next Saturday” is a different day. The same run caught a store contract filling itself from a law office's call; contracts now skip conversations that are not about their use case.
Measured on a laptop CPU
Sources
- faster-whisper · SYSTRAN, GitHub
- Llama 3.2 on Ollama · Ollama