The shape of this problem
Too much arrives to check it all, so you sample
The problem
Call-center teams needed searchable Arabic transcripts across dialects. Recorded calls that cannot be searched are closer to a liability than an asset - they must be retained and produced on request, but nothing can be learned from them. Arabic sharpens the problem: callers speak in dialect while most systems are built for Modern Standard Arabic, so accuracy varies with who happens to be calling, which is the worst possible property for a compliance record.
What we built
We delivered speech-to-text processing that converted conversations into structured text suitable for search and downstream analysis.
What moved
| Measure | Before | After |
|---|---|---|
| Word error rate on Gulf-dialect calls | 31% | 12.7% |
| Spread between best- and worst-performing dialect groups | 19 points | 5 |
| Call archive searchable | no prior measure | all of it, previously none |
Word error rate on Gulf-dialect calls
Spread between best- and worst-performing dialect groups
Figures are from the practice's own delivery records for the engagement named, measured against the process that preceded it.
What it turned on
Handling dialect variation is what turned the archive from stored audio into a queryable record - uneven accuracy across callers would have made any aggregate drawn from it misleading. Relevant to any regulated operation retaining Arabic voice records it currently cannot read.
- Service line
- Service & Conversation AI
Start here
Which of the four is yours?
Tell us the documents and the monthly volume and we will send the two closest records, with the figures behind each and what it would take to repeat them on your material.
You get a reply within one working day, from the engineer who would do the work.