Microsoft introduces new intelligent CQD media quality Classifiers

— This blog is part of a series of technical articles on Microsoft 365 products in general and Microsoft Teams specifically —

As we know, the Call Quality Dashboard (CQD) for Microsoft Teams provides insights into call quality. It does so by using basic classification on media streams to identify “Good” and “Poor” audio, video and screensharing. Microsoft recently introduced new, so called, intelligent media quality classifiers in CQD. These new classifiers use Machine Learning (ML) algorithms to analyze call telemetry and identify potential issues in call quality. So, what do they do and how do they differ from the traditional stream qualifiers? And does this mean you can now do without other tooling?

In CQD, Microsoft traditionally classifies streams for audio, video, and video-based screen sharing (VBSS) based on network and video metrics. Traditional stream classification refers to the method used to categorize media streams based on key quality metrics. Streams are classified as Good, Poor, or Unclassified depending on the values of these metrics. As an example an audio stream might be marked as Poor if certain conditions, such as high packet loss rate or excessive jitter, are met.

The new intelligent media quality classifiers, however, take a broader view. They consider multiple factors to assess user experience and identify root causes of quality degradation. These classifiers offer a more advanced analysis compared to traditional stream classification. It should help IT admins pinpoint specific problem areas, such as network, compute device, or input device issues, more easily.

There are two levels of classification for the new Machine Learning driven classifiers: higher-level and lower-level.

Higher-level classifiers predict if audio, video, or VBSS aren’t functioning properly. Lower-level classifiers identify if the issue is related to the network, compute device, or input device. In generalized terms: The higher level classifiers tell you what is likely impacted while the lower level classifiers tell you what the likely cause is.

All dimensions and measurements available to you (both old & new) can be found here. To filter out the new ones is a bit harder but luckily Microsoft also updated the QER Power Bi Template and included a handy guide with it of the new measurements:

Using CQD  you will find the new dimensions and measurements.

For example, using CQD you can search for “Problem” and you will see most of the new dimensions (see screenshot).  In this case we want to know how many of the inbound streams in calls had inbound network problems vs how many didn’t or weren’t classified:

If you now combine these new meassurements with a dimension like month/year and add the total Calls dimension as a line chart, you can create graphs like the following:  

Why are there so many unclassified (unknown) streams? Well because the classification can only take place under certain conditions (see the documentation for those conditions) and many of the streams apparently didn’t meet those conditions.

These new Intelligent classifiers use machine learning algorithms to analyze call data and identify potential issues affecting call quality. They look at various factors, such as network performance, device performance, and user feedback, to predict if a call was good or poor.

However there is an important sentence which Microsoft mentions in its article which we need to consider:

Though driven by intelligent mechanisms, it might still be guessing, but isn’t predicting also a form of guesswork? Microsoft uses the term predicting in this context and that is right! It’s also what sets it apart from their existing classifiers and therefore means it can differ from them.

CQD still lacks both a consolidated and deep-dive perspective for research and analysis. Particularly when the objective is to examine data on an individual user or a specific group of people, and even more so when additional contextual data is required. If you are interested in how TrueDEM is doing all of this and allowing you to do deep dive research and analysis then take a look here or contact us.

We would love to hear from you!