Comments on: What Does Machine Learning Have to do with MOS Scores? https://bloggeek.me/machine-learning-and-mos-scores/ The leading authority on WebRTC Sat, 02 Jul 2022 12:49:51 +0000 hourly 1 By: Jaro https://bloggeek.me/machine-learning-and-mos-scores/#comment-119582 Sun, 02 Jun 2019 09:29:35 +0000 https://bloggeek.me/?p=12785#comment-119582 Hmm – “opinion” – it is about the interaction between 2 humans and the technical quality use to be only one of the factors causes the “positive” or “negative” feeling about the dialogue. So – we use different approach in the O2 CZ contact center. The user has an option to “reward” the operator by thumb up or down when the session finishes. Cca 12% of the conversations are “rewarded” – most of it positive. But if we include in the game AI (we are using .NET ML) – we are able to guess the “thumb” with accuracy of 90-95%. Because we feed the AI not only with technical parameters, but with “all we know” about the session. It means time, duration, scope, agent name, customer past experience and some of technical data mentioned in Tsahi’s article. This approach can improve the performance of call center significantly – we are able to find “not rewarded”, but potentially negative sessions and anticipate the negative reaction according past (computed) experience.

]]> By: Tsahi Levent-Levi https://bloggeek.me/machine-learning-and-mos-scores/#comment-119295 Wed, 19 Dec 2018 09:12:55 +0000 https://bloggeek.me/?p=12785#comment-119295 In reply to Ilya.

Ilya,

I have no clue 🙂

Which is why I hate MOS so much as a predictor of audio quality. There are just too many factors to place into a single number between 1-5.

]]> By: Ilya https://bloggeek.me/machine-learning-and-mos-scores/#comment-119294 Wed, 19 Dec 2018 09:06:54 +0000 https://bloggeek.me/?p=12785#comment-119294 ” the device magically “improves” the audio simply by reducing the noise.”

In other words, it’s some type of noise canceling headset where ML is used for providing typical noise patterns (‘office’, ‘street’, ‘subway’ etc) instead of traditional noise canceling techniques. Does noise reducing always means higher MOS? MOS, for example, depends on language. If a UK caller calls a callee in India, which call leg defines final MOS – UK or Indian? How to normalize MOS in a network for different languages (accents, pronunciations etc)?

]]>