01 / The problem
What the system needs to solve
Call centers need every call reviewed for compliance, objections and coaching. Manual QA covers only a small sample, and cloud transcription sends sensitive customer audio outside the organization.
02 / The approach
How the pieces fit together
- Transcribe with NVIDIA Parakeet TDT v3 for word-level timestamps.
- Separate agent and customer with NVIDIA Sortformer diarization, then align every word to its speaker.
- Compute conversation metrics: talk ratio, interruptions, dead air, silence and words per minute.
- Score each call with Qwen3 8B through Ollama for compliance, sentiment, objections and coaching feedback.
03 / Key features
What it does
- Dashboard with synchronized audio playback, searchable transcripts and QA scorecards
- Transcript confidence checks and second-pass verification of uncertain segments
- Risk tagging for numbers and compliance-related words
- Evidence validation that ties AI findings to the transcript
- GPU memory monitoring with automatic recovery from CUDA out-of-memory errors
04 / Results
What it delivered
- ≈38× real-time transcription speed on an NVIDIA RTX 5070 Ti
- ≈16× real-time sustained end-to-end processing throughput
- Customer call audio never leaves the organization's own hardware
05 / The tradeoffs
Engineering is a set of choices
Running speech and language models locally keeps call audio in-house and removes per-minute API costs, but throughput is bounded by one GPU's memory. Concurrency and out-of-memory recovery have to be engineered rather than bought.
06 / Validation considerations
What to test before relying on it
Concurrency tests under load, GPU memory monitoring, and evidence checks that link every AI finding back to a timestamped transcript segment.