Large-Scale Speaker Verification System for Viettel Call Center
Published:
Duration: Nov 2021 – Sep 2023
Organization: Viettel Big Data Analytics Center, Viettel Telecom · Hanoi, Vietnam
Overview
Led the end-to-end development and production deployment of a large-scale speaker verification system for Viettel’s call center — Vietnam’s largest telecom operator. The system replaced traditional OTP and security-question authentication with voice biometrics, serving millions of customers daily.
Key Contributions
- Designed and trained speaker embedding models on large-scale Vietnamese call-center audio data
- Built the full inference pipeline and deployed with Triton Inference Server for high-throughput, low-latency serving
- Reduced manual verification workload by >50% across Viettel’s call center operations
- Achieved ≈99% accuracy on internal Vietnamese speaker datasets, outperforming existing baselines
- Ranked Top 20 in the VoxSRC 2022 international competition (beginning phase)
- Integrated data pipeline using Kafka for real-time audio streaming from call-center infrastructure
Technologies
Python PyTorch Triton Inference Server Docker Kafka Speaker Verification ASR VAD NLP