Telco Contact-Center Live Transcription Blueprint

Problem/Feature Description

A Romanian telecommunications company runs a contact center handling approximately 200 simultaneous customer service calls at peak. All calls arrive from the public telephone network at 8 kHz, and the company needs accurate real-time transcriptions delivered to agent assist dashboards with words appearing as the caller speaks. Agents are not willing to wait more than two seconds for text to appear on-screen after a caller finishes a phrase.

The company's language mix is exclusively Romanian. They operate on-premises with a cluster of NVIDIA data-center GPUs (budget for up to four cards) and have no tolerance for vendor-locked cloud APIs, having had a contract dispute with a cloud ASR provider the previous year. The legal team requires that all model licenses permit commercial use without per-call royalties. The company has heard about several different ASR models and frameworks but has no strong prior commitment to any specific one.

Your task is to produce a complete architecture blueprint for this system. The blueprint should cover the full stack from audio ingest through transcription delivery: which model to use and why, how audio is transported and normalized, how the streaming decode is structured, how inference is served across GPUs, how hardware is sized, what runtime and precision settings to use, what supporting processing components are needed, and how the system will be validated before going to production.

Output Specification

Write the blueprint as a single Markdown file named blueprint.md. The document should be self-contained and ready to hand to the engineering team. Include a concrete sizing calculation with explicit numbers the team can revisit if concurrency assumptions change. Include an evaluation plan with specific metrics and benchmark datasets appropriate for this language and domain. Make deliberate picks and name the alternatives you considered but rejected, with a reason for each rejection.