Three Misconceptions About Speech Analytics, or What Every Marketer Should Know in 2020

[ashammi228]

Content
Misconception 1: Without verbatim transcription, speech analytics is useless.
Misconception 2: Every conversation is unique - a machine algorithm will not understand the essence
Misconception 3. It is enough to listen to 10% of calls to find out the average temperature in the hospital
Conclusion

Semyon Suslin, Speech Analytics Product Manager

Speech analytics has let the genie out of the bottle. Just telegram brazil amateur imagine: you have access to 100% of customer conversations with managers. Not everyone is ready to apply this knowledge, but it is the future: if there is data, businesses will use it. We previously wrote about what speech analytics is capable of . In this article, I will tell you what myths and misconceptions prevent marketers and sales managers from successfully using new data about customers and deals.

Misconception 1: Without verbatim transcription, speech analytics is useless.
Speech analytics is a technology based on machine learning methods that translates human speech into text format. In recent years, the error rate for a number of technologies does not exceed 4-5%. And although decoding Russian speech is complicated by the large number of phrases in a sentence, the recognition accuracy reaches 80%. That is, 8 out of 10 words will be recognized and available for searching for the desired dialogue, tagging, training the neural network taking into account the context.

We have heard from clients more than once: speech analytics does not provide a verbatim transcript, not all words in sentences agree, how to read this? But, attention, a question: does the employee who listens to calls translate them into text? No! He tags, makes notes, finds key words. If you force him to transcribe what he heard and type the text, you will have to pay for a whole company of specialists 24/7.

Speech analytics ≠ transcription

Speech analytics, just like a person, marks calls by the presence of mandatory and/or stop words. Only a person manages to do this for 5-10% of calls, and the program - for 100% of conversations. That's the only difference.


Excerpts from transcript of conversations between client and manager
It is also wrong to think that the intonation and complexity of the great and mighty Russian language are only understandable by humans. The accuracy of human decoding of recorded spontaneous speech reaches 99%. The accuracy of speech recognition technology for subsequent analysis by keywords, as we have already noted above, is 80%. This figure is successfully increased by training the algorithm, supplementing the dictionaries with terms and specific vocabulary. Thus, this year we are adding three new industry dictionaries to our speech analytics system: "Automotive topics", "Medicine", "Real estate". Moreover, in the latter topic, the tagging accuracy for a number of projects already reaches 96%.



Human

Artificial intelligence

Transcript accuracy (for keyword analysis)

99%

80%

Share of calls listened to

10%

100%

If we take the entire array of records as a unit, we get a potential data “coverage” of 9.9% for a person, and 80% for a machine. The difference is almost 10 times!

There is no need to wait for ideal transcripts. It is enough to take into work an array of data in the broken language of artificial intelligence and in a few clicks find in it the calls that a specialist needs to listen to. And if before connecting speech analytics the employee analyzed random 5% of calls, then after he can do the same amount of work, but only on those calls that require special attention. For example, find calls in which the operator and/or client used a certain word.


One of our clients listened to calls tagged "negative" and discovered an interesting pattern. It turned out that some customers, when receiving goods from a courier, were unhappy with the extra charge for lifting to a floor without an elevator or delivery outside the Moscow Ring Road. This point was added to the script. Now, at the end of the conversation, the operator immediately announces the final cost of the order and delivery. Compliance with this point is monitored for all calls using automatic tagging. The source of negativity has been completely eliminated.

Important to know

Speech analytics is not about transcribing voice into text. It is about fast access to a large volume of data that businesses lack to make the right management decisions and increase sales.

It would seem that the task of analyzing calls is simplified: it is enough to make a list of marker words and tag calls by them. And again the question: can you name 10-20 phrases that characterize a call as high-quality and targeted?

Increase sales with the UIS communications platform
A reliable cloud telephony operator: own number capacity and technical support No. 1 on the market.

Manage communications, control employees and automate the sales department.

Get a consultation

Misconception 2: Every conversation is unique - a machine algorithm will not understand the essence
Marketers and sales managers should work together to determine the parameters that characterize a high-quality, targeted call. And this is not a whim, but a vital necessity given the trend to work across the entire sales funnel. This exercise is also useful for those who are not yet thinking about implementing speech analytics.

Our experience shows that about 40% of companies do not have such criteria, or the opinions on this matter differ greatly between marketers and sales managers. This is where this eternal confrontation comes from:


Meanwhile, a telephone conversation that most likely leads to a sale always has a number of key points. For example, these could be questions about delivery for a furniture store, or the mention of a specific doctor's name for a clinic. The more of them, the higher the probability of a sale. To find these triggers, you need to listen to at least 10-20 conversations, including separately those that led to a sale and those that ended in nothing.


You can argue for a long time that the machine does not understand the nuances of our speech. But the facts say otherwise: when analyzing calls in any particular topic, the machine learning algorithm needs 15-20 trigger words to tag calls. 15-20, Karl!