Why Persian Speech Recognition Stumbles on Numbers in the Contact Center
On a phone call, one wrong digit in a tracking code makes the whole conversation useless. Lessons from the ParsCallScribe research on numbers, data and inverse text normalization.

Saeian Ertebat (Sena Telecom)
When we talk about speech recognition accuracy we usually report one number: word error rate (WER). It is useful for comparing models, but in a contact center it hides what matters. A customer reading out a case number does not care how many other words the model got right. If one digit is wrong, the call fails.
Why numbers are hard in Persian speech
In conversational Persian the same number is spoken in several ways: as a full number, digit by digit, in groups, or as a mix. Narrow-band telephone audio makes it worse. Most models also write numbers out as words, while downstream systems need digits.
What the research showed
In the ParsCallScribe doctoral research we trained a fairly small model (QuartzNet, 19M parameters) on a 512-hour corpus from three sources: public data, survey recordings and real contact-center calls. Online augmentation raised the effective volume to about 2,000 hours.
The result had two sides:
- Overall WER on 4,000 test conversations was 18.4%.
- On 1,000 number-centric utterances, after adding an inverse text normalization (ITN) layer, WER was 2.7%.
So for the part of the call that carries the most business value, capturing numbers and codes, reliable accuracy is possible even while overall conversational accuracy still has room to improve.
Three practical lessons
Take the metric from the business. Before choosing a model, ask which part of the conversation drives the decision. In a contact center it is usually numbers and names.
Real telephone data has no substitute. Even a few dozen hours of real calls, next to public data, prepare a model for real conditions.
Take post-processing seriously. ITN is a small part of the system, but in this research it closed the gap between output you can read and output you can use.
A bigger model is not always the answer. Sometimes a small model, the right data and a simple but precise layer solve the real business problem better.