Speech Tag Training Reference Guide

Overview

Speech Recognition provides tools for quickly understanding key elements that may or may not be present in conversations agents have with customers. With tools such as full-text transcription, contact centers can quickly perform searches for specific words or phrases or have the system identify these on an ongoing basis as tags.
Emotion detection can identify conversations where customers are unhappy, or even more specifically where a customer’s mood was improved or made worse by interacting with our agent.

This guide is meant as a reference while a consultant works with the contact center to establish the most effective speech tags for the business.

Audience

This document is intended for those who have purchased the Speech Recognition module and are responsible for setting up speech tags within the Eleveo Workforce Optimization solution.

Key Terms

Speech Recognition - Elevēo's Speech Recognition solution includes: Transcription Services, Emotion Detection, and Phrase Detection

Auto Quality Management (AQM) - A tool that can automatically evaluate a call-based conversation based on created rules and provide a score on all transcribed conversations.

Transcription Services - Speech-to-text for recorded calls with punctuation and number formatting

Emotion Detection – Ability to identify positive, neutral, negative emotions within a conversation. Additionally, the system may note when moods improve or worsen.

Phrase Detection – A process by which a collection of words and phrases are labeled to help businesses meet important objectives.

Redaction of Sensitive Data - Ability to redact sensitive numbers and replace with ##### instead of displaying numbers

Full-Text Search - Ability to search conversations that include or exclude specific terms, names, or text

Speech Recognition

The Speech Recognition engine generates transcriptions of conversations that may be viewed in Conversation Explorer.

If a user/administrator has also set up pre-defined Speech Tags, the engine will also review these during the transcription process, tagging the moment in the conversation the phrase or word was detected with an icon.

Redaction

Eleveo takes personal data seriously and supports the redaction of sensitive data from transcriptions. The redaction of credit card data and other sensitive numbers is enabled by default when the solution is first configured.
Any redacted information is shown as #### (hashtags) in the transcription so that viewers can clearly see that something was hidden.

Emotion Detection

The average emotion at the end of the conversation is extracted from the transcription and then displayed within the Conversation Explorer Details Pane.

Speech engine emotion detection feature uses a synthesis of acoustic features and word sentiment scores to determine if a given utterance is Positive, Negative, Neutral, Improving or Worsening.

Phrase Detection – Speech Tags

How Does it Work (in Brief)

To find speech phrases that have been pre-defined by the user, the system performs several steps.

The list of speech phrases is compared to the transcription of audio recordings for each channel. To do this the system:

Splits each speech phrase into a list of words,
Compares the transcription of the recording with the first word from the list of Speech Phrases (Step a) until the words are matched. When a match happens, the system then attempts to match the whole speech phrase (Step a) within the transcription of the same channel. When there is a match, the system saves the speech phrase occurrence to the database.
Found phrases are displayed within the Conversation Explorer.

How to Prepare Speech Tags and Phrases

Speech Recognition searches for targeted phrases which are user defined and grouped according to categories referred to as Speech Tags. Optimally, the groupings (Tags) should be created with the following in mind:

As the system searches for an exact match, phrases should be short. This will ensure that the search terms are sufficiently unique to be found.
If the system is having trouble catching all or most phrases of importance, consider shortening the phrase further. For example, is there an “anchor word” that could be searched for? (e.g. instead of searching for “may I have your social security number” search for “social security number” alone)?
Consider using different ways of saying the same thing (e.g. “Thank you for calling” and “Thanks for calling”)
As a phrase will only be captured by one Tag, do not use the same phrase in multiple Speech Tags.

Example Speech Tags

Negative Sentiment	Positive Sentiment	Customer Verification	Customer Greeting
I can't believe this	That's incredible	Social Security Number	Thank you for calling
This is ridiculous	I really appreciate that	Date of Birth	Thanks for calling
I'm unhappy about	That's so amazing	Address	Hello, my name
File a complaint	Thank you so much	Phone Number	Hello, this is
So frustrating	You are the best	Your Social	Good morning
Very frustrating	This is excellent	SSN	Good afternoon
Pissing me off	You just made my day	Last four digits	Good Afternoon
Aggravation	That's such a relief	Account number	This is Customer Service
Frustration	You're the best	Reference number	Welcome to ABC
So aggravating	That's a relief	Contract number	How can I help you
Very aggravating	You're a life saver	Last payment made	How may I help you
Speak to a manager	That made my day	Email address	How may I assist
Speak to a supervisor	You're amazing		How are you today
Speak to your supervisor	You are amazing		Hi, How are
Talk to a manager	So grateful		Help you
Talk to a supervisor
Transfer me to a manager
Transfer me to a supervisor

Speech Tag Examples cont.

Upsell or Close	Fraud	Product Knowledge Gap Phrases	Closing statement
Buy today	Write off	Need to check	Anything else
Can I activate	Failed investment	Need to verify	Appreciate your
Discount off	Off the books	Let me check	Enjoyed Talking
Agree to	Nobody will find out	I don’t know	Once Again My Name
Have you considered	Grey area	I can't help you	Glad I could help
Set a time	Illegal	I'm not sure	Any further questions
We can deliver	Do not volunteer info	Let me call you back	Brief survey
Actually offer	Not ethical	Let me look into that	Happy to help
Handle that	They owe it to me	I guess
Like to go with	Backdate	I'll get back to you
	No inspection	Find that out for you
	Pull earning forward	I'm new
	Special fees	Need to ask someone
	Friendly payments	Put you on hold
	Do not share
	Expected to announce
	Deserve to get paid
	Breach
	Top secret
	Cook me up
	Cover up

FAQs

How many syllables should my word or phrase be to be most successful?

We suggest that each Speech Phrase be short. (one to 3 words) This will ensure that the search terms are sufficiently unique to be found.

Only exact matches will be found.

Is there a maximum duration for a word or phrase to be detected?

No, but remember that the search is looking for exact matches on the transcribed text. The longer the text, the less likely an exact match will be located.

Is there a maximum call length that can be scanned by the Speech engine?

No, as the search engine is looking at the transcribed text, there is no max call length.

When I add a new Word or Phrase to my Speech Tag, does that apply to interactions already recorded and saved to my system?

Phrase detection occurs once a transcription has been processed and saved.

How quickly is a call transcribed?

At a minimum, transcription should be set to occur 1 hour after call completion.

Can we re-scan calls or scan calls in the past?

Re-scan is currently not supported but, we are working actively to enable this feature.

Can I use special characters (like commas, periods, and question marks) when creating Speech Phrases?

Yes, the following special characters are supported ! - + . ^ : , ?

Can numbers be found by speech detection?

Yes. Speech tags support alphanumeric characters. Speech tags support both numbers and words. (e.g. 20 min / twenty minutes).

Reminder that sensitive numbers are redacted and changed to ###. In this case, the search function or speech detection will not locate the number.

Are phrases case sensitive?

No, the algorithm is not case-sensitive.

How is emotion determined?

The speech engine emotion detection feature uses a synthesis of acoustic features and word sentiment scores to determine. Resulting in a Positive, Improving, Neutral, Worsening, or Negative response emotion tag.

How does the algorithm deal with the use of contraction words, such as it's vs it is?

At the moment, the algorithm supports exact match only. It is NOT equal to it is. We recommend that multiple phrases be added to cover all variations that may be used

How much of the call does the system use to recognize which language to transcribe the call into?

The system uses the first 20 seconds of the call to identify the language. If multiple languages are used within the first twenty seconds (e.g., English and Spanish), the system will default to English.