How Real-Time Speech Analytics Is Improving Agent Quality Scoring for a Contact Center BPO

The Challenge

Manual quality review covered only a small sample of each agent's calls, so scores often reflected which calls were picked rather than how the agent actually performed.

The existing approach made it difficult to:

• Score enough calls per agent for coaching decisions and client reports to be statistically meaningful.
• Apply each client's scorecard consistently when different evaluators read the same criteria differently.
• Catch missed compliance disclosures, such as recording notices or payment-card language, before they became client escalations.
• Step into difficult calls while the customer was still on the line instead of reviewing them days later.
• Give agents specific, timestamped feedback rather than general comments.

From a technical perspective, contact center audio is hard to work with. Calls arrive as 8 kHz telephony streams with crosstalk, hold music and background noise, agents and customers speak with a wide range of accents, and each client brings its own product names and scripts. The pipeline had to separate agent and customer speech, keep transcription latency low enough for live alerts, redact card numbers and personal data before storage, and run inside the BPO's existing telephony and workforce tools without disrupting live operations.

The Solution

InnovationM designed a streaming pipeline that taps call audio from the telephony platform, converts it to text with speaker labels, and passes each conversation through scoring models built around the BPO's actual QA forms. The team began by collecting a labeled set of calls scored by senior evaluators and used it to benchmark speech engines and tune every scoring rule. The solution included:

• Streaming speech-to-text: a real-time ASR layer with per-client custom vocabularies for product names, plan codes and brand terms, benchmarked on the BPO's own recordings for word error rate.
• Speaker diarization and channel separation: stereo channel handling and diarization so every phrase is attributed to the agent or the customer.
• PII and PCI redaction: detection and masking of card numbers, account details and personal identifiers in both transcripts and stored audio.
• Scorecard engine: a configurable mapping from each client's QA form to detection rules and LLM-assisted evaluators for greeting, verification, empathy, resolution and closing criteria, each with evidence quotes and timestamps.
• Live supervisor alerts: real-time flags for missed disclosures, escalation language and negative sentiment trends, pushed to a supervisor dashboard during the call.
• Calibration and coaching workspace: a review screen where evaluators accept or override automated scores, with overrides fed back to improve the models, plus agent-level trend reports connected through InnovationM's data engineering practice.

The key focus was not simply transcribing more calls, but making every score explainable to the agent, the supervisor and the client who reads the report.

The Impact

Quality review moved from a sampled, after-the-fact exercise to a continuous process where every call is scored and humans spend their time on calibration and coaching. The solution supported:

• Higher share of calls scored per agent.
• Closer agreement between automated and evaluator scores after calibration.
• Fewer missed compliance disclosures reaching client audits.
• Faster supervisor intervention on escalating calls.
• Timestamped, evidence-based coaching notes for every agent.

By bringing together streaming transcription, redaction, scorecard-aware evaluation and a calibration loop, the BPO can show each client a complete, consistent picture of quality instead of a sample.