- Year
- 2017 – 2026
- Organization
- Islamic Azad University, North Tehran Branch
- Industry
- Contact centers & speech research
- Role
- Doctoral researcher
Personal vs. team contribution
Personal contribution
Corpus design, model training and evaluation, and the doctoral thesis.
Team & organization
Supervised by Dr. M. Manthouri.
The challenge
Persian telephone speech is narrow-band, conversational and full of numbers (phone numbers, tracking codes, amounts); general models make many errors on it.
The solution
A 512-hour corpus from three sources (Common Voice, survey recordings and real contact-center calls), online augmentation to roughly 2,000 effective hours, a compact QuartzNet 15×5 model with 19M parameters, and an inverse text normalization (ITN) layer for numbers.
Architecture
- 8 kHz telephone audio
- Online augmentation
- QuartzNet 15×5 acoustic model
- Decoding
- Inverse text normalization (ITN)
Technology stack
- Python
- PyTorch
- NVIDIA NeMo
- QuartzNet 15×5
Impact
18.4%
WER on 4,000 test conversations
Research result
2.7%
WER on 1,000 number-centric utterances (with ITN)
Research result
0.1
real-time factor (RTF)
Research result
Lessons learned
In a contact center, getting numbers right matters more than overall accuracy. A small model with the right data and an ITN layer reached 2.7% WER on number-centric calls, while overall conversational WER was 18.4%.