97% classification accuracy on test data, with 5M support calls a month scored in production

A deep learning engine that measures audio quality on two-way customer service calls, pinpoints the source of each issue, and maps call quality by geography and carrier.

Outcome

5M calls processed a month, 97% overall classification accuracy, processing time down 50%, support hours down about 60%, customer satisfaction up 25%, and an estimated $5M saved a year.

See the results

Talk to an amp'd specialist

At a glance

Client & industry

Global technology leader, with two-way customer service calls between representatives and end customers

Engagement

Custom audio quality analysis on Vertex AI, productionized along with MLOps

Scale

5M calls a month, with 5 to 7% affected by audio quality issues

Business owner

The client's availability and reliability team

Maps to below services

AI & ML Engineering, MLOps & Scalable ML Platforms

Where it started

Audio quality issues affecting 5 to 7% of customer service calls, with no reliable way to catch or trace them.

Business problem

Each recorded call was evaluated manually and given one overall rating based on the audio issues observed. Evaluating that many call files by hand was cumbersome and time-consuming, and support agents needed real-time help to reduce resolution time.

Technical barriers

Existing tools used RMSE and spectrograms to visualize and analyze signals, so echo, lag, low volume, and background noise went unrecognized. Issues were never captured or scored by type, root cause analysis and proactive monitoring of problem calls were missing, and call quality was never tracked at regular intervals by geography, carrier, time of day, or day of week.

Agivant's role

Deep learning model development, audio processing pipeline engineering, scoring framework design, and call insights dashboards

What Agivant built

Agivant built and implemented a complete audio quality solution on Vertex AI. Call recordings are copied from storage into a Dataflow pipeline, split into 30-second segments, and scored segment by segment through model inference. Speaker diarization distinguishes and labels each speaker, a weighted scoring framework combines attribute scores into one call score, and results land in BigQuery for dashboards and anomaly detection. Value in days and weeks.

Multi-attribute quality scoring

Deep learning models measure multiple audio quality attributes on two-way calls, detecting echo, lag, low volume, and noise.

Speaker diarization

Distinguishes and labels each speaker in the recording, segmenting the content by speaker so the source of an issue during a call is clear.

Weighted scoring framework

Combines attribute scores into one customized, weighted quality score per call.

Call insights dashboards

Maps regular performance, with quality distribution by geography, carrier, and other call metadata, and highlights anomalies.

Anomaly detection

Quantitative measures and statistical analysis flag deviations from normal patterns, such as abnormal CSAT scores or average handling time alerts.

Parallel processing pipeline

Dataflow processes call files in large-scale parallel runs, with copies purged after processing and accessible only to the pipeline.

Audio scoring outputs

Each call produces an audio scoring summary and segment-level details, read through the BigQuery Storage Read API.

ML lifecycle on Vertex AI

Model training, deployment, registry, serving, and monitoring all run on Vertex AI.

How it is built

ML platform	Processing	Storage and results	Data access	Orchestration	Controls
Vertex AI training, deployment, registry, serving, and monitoring	Cloud Dataflow ML	Cloud Storage, BigQuery	BigQuery Storage Read API	Cloud Composer, Workflows, Cloud Scheduler	Cloud IAM, Cloud Monitoring, Cloud Logging, file copies accessible only to the Dataflow pipeline

Before

Manual call-by-call review, one overall rating, echo and lag unrecognized, and no pattern tracking.

After

Every call segmented, scored per attribute and per speaker, rolled into a weighted score, and mapped by geography and carrier.

Proof in production

A call recording is copied into the Dataflow pipeline and split into 30-second segments. Vertex AI predicts quality on each segment, diarization shows who was speaking when the issue occurred, and the scores land in BigQuery. Dashboards map quality by geography and carrier, anomaly detection flags unusual CSAT or handling time patterns, and the copy is purged once processing finishes.

Measured results

Call quality issues are now identified automatically and traced to their source, with fewer agent hours per call.

5M

Calls processed a month

97%

Overall classification accuracy on test data, using 32 audio metadata attributes

80%+

Accuracy on 5 audio quality attributes

50% down

Processing time

~60% down

Support hours

+25%

Customer satisfaction score

Est. $5M

Annual cost savings

The repeatable pattern

Reusable asset

The audio quality measurement framework is an approved Agivant solution on Google Cloud, and it runs on the Vertex AI MLOps platform that keeps its models accurate. Every build amplifies the next.

Challenges overcome

Recognizing issues that signal-level tools missed, classifying quality across 32 audio metadata attributes in high-dimensional space, and processing call files at scale while keeping them accessible only to the pipeline.

Enterprise movement

Foundational

Operational

Path to autonomous