Back to Projects

CDO Speech Services

Enterprise AI Solution

Unified speech and translation services abstracted into a single API and Python client library.

Overview

An enterprise speech services platform that abstracts complex speech-to-text, text-to-speech, and translation capabilities into a simple, unified API. The platform includes a Python client library for easy integration, enabling teams across AT&T to add voice capabilities to their applications.

Challenges

  • Integrating multiple speech engines with different APIs and capabilities
  • Supporting real-time streaming for live transcription use cases
  • Handling diverse audio formats and quality levels from various sources
  • Providing accurate transcription for telecom-specific terminology
  • Ensuring low latency for interactive voice applications

Solutions

  • Built abstraction layer over NVIDIA Riva with fallback to cloud providers
  • Implemented WebSocket-based streaming API for real-time transcription
  • Created audio preprocessing pipeline for format normalization and noise reduction
  • Fine-tuned acoustic models on AT&T call center recordings and technical vocabulary
  • Deployed on GPU-enabled Kubernetes clusters with auto-scaling based on demand

Key Results

Achieved 95% word error rate accuracy on telecom-specific conversations
Reduced integration time for new applications from months to days
Processing 100,000+ minutes of audio monthly across 15+ applications
Enabled real-time captioning for internal meetings and customer calls
Saved $500K annually by consolidating multiple speech service contracts

Technologies

NVIDIA RivaKubernetesPythonFastAPIgRPCWebSockets