Best practices for Phonetics Boosting

To get the most from your Phonetics Boosting projects, first understand the main causes for poor recognition of terms and phrases, and then follow the ten guidelines when adding them.

Limit boosting to only business-meaningful terms

Avoid the tendency to correct every term that is transcribed incorrectly. Make sure that the terms and phrases you decide to boost are business-relevant or business-meaningful. Product names, campaign names, new offerings, and competitor names, are examples of business-meaningful terms to boost.

Example: If you are in the insurance industry, you may want to boost phrases such as full coverage or for life. There is no true value in boosting everyday language with no impact such as plastic or drawer.

The boosted terms and phrases must go hand-in-hand with your Speech Analytics or Real-Time Speech Analytics categories to improve your categorization accuracy.
Boost phrases as opposed to individual terms

To get the most benefit, boost phrases as opposed to single terms. By providing the context for key terms, phrases with three terms (trigrams) or two terms (bigrams), improve the recognition rate and accuracy of your newly boosted model.

Example: Phrases such as purchase an iphone or buy an iphone, rather than just the term “iphone".

Special characters such as & ! @ # $ are not valid in the language model. To boost a phrase with special characters such as S & P, you would add it as S and P.
Remember that knockout is not equivalent to removal

The knockout option in Phonetics Boosting, while assigning the term the lowest probability, still retains the term or phrase in the language model.

What happens when you knock out a term or phrase?
- There is expected to be a significant decrease in a term or phrase after it is set to knockout on the probability band.
- If Speech Analytics does not find a similar-sounding term or a phrase with higher probability, term and phrases with knockout priority can still appear in transcripts.
- To completely remove terms from the language model, contact Verint Product House.
Example: You may want to knock out the phrase Verint wife to prevent it from being transcribed instead of Verint Life.
Suppress competing terms and phrases

Lower the priority of terms and phrases that sound similar to other terms and phrases, but are not meaningful to your business.

Example: If the system transcribes Verint as inherit, the recommendation is to reduce the priority of the term inherit, and boost the priority of the term Verint.
Boost frequently used phrases

Add phrases frequently repeated in conversations to improve their recognition. These phrases can be standard opening and closing phrases used by employees as per scripts, or any other phrase that recurs in employee-customer conversations frequently enough to justify boosting.

Use the comma as the delimiter to add several phrases at once. To cover the wide range of variations, you can also include several variations of the same phrase, separated by commas.

Example: Hello how may I help you, Hello how may I assist you, is there anything else I can do for you, anything else I can assist you with today

The comma as the delimiter is the only punctuation considered valid in the project.

All other punctuation such as question marks and periods are invalid. For example, how may I help you? is invalid because of the ? at the end of the phrase.
Use asterisk (*) as wildcard in phrases

Use the asterisk (*) as a wildcard to replace any term that can vary in the actual conversations. For example, employee names, company and department names, and common terms that are used interchangeably.

Examples:
- Hello thank you for calling my name is *
- how can I * you
  
  where:
  
  * can be either assist or help
  
  The asterisk is valid only within or at the end of the phrase. There must be a space between the asterisk and the term before it.
Boost the linguistic connection with similar terms

If there are multiple business-meaningful terms that sound the same but have different meanings, use them within expressions that demonstrate the linguistic connection.

Example: Boost both meanings of the word check in the relevant context, by using expressions such as pay by check and check this for you.
Use the probability band to track your changes

Use the scale bar to adjust the detection probability of terms and phrases. Always check the old and new arrows to determine the current probability of a term or phrase, and its new probability after boosting. Ensure that the location of new arrow matches your requirements.

Phonetics Boosting automatically suggests Band Adjustments, typically, either Unchanged for terms that do not need boosting, or Medium for terms that need boosting. The best practice is to retain the automatic suggestions.

If you want to use Advanced options, such as Knockout, Low, and High, make sure you understand the impact of the selected option:
- Knockout: The term will rarely be transcribed.
- Low: Low probability that the term will be transcribed.
- High: Very high probability that the term will be transcribed.
To improve associated results of a term or phrase previously boosted to Medium, change the Band Adjustment to High.

Boosting a term or a phrase gives it a Medium ranking by default (or Unchanged for terms with high probability). If you boost a term that was already ranked as High to Medium, you will be lowering the probability.
Add pronunciations

If a term has several pronunciations, add all the pronunciation variants to boost recognition.
- Derive pronunciations for complex terms
  
  Make sure you add pronunciations according to how the term or phrase is spoken. If the term you are adding is not in the dictionary, instead of creating your own pronunciation, break it into logical parts and then derive the pronunciation from similar terms in the dictionary.
  
  Example: Assuming the term Telcomind is not included in the dictionary and you want to add it, you can search for and create it from the existing dictionary terms Telco and mind.
- Add multiple pronunciations
  
  If you know that the term or phrase has more than one pronunciation, make sure you add all the known pronunciations for that term or phrase. For example, the letter o in the term Boston can be pronounced as the o in orange, and as the a in bar.
  
  Do not have multiple spelling representations for the same term or phrase, as they reduce detection accuracy.
  
  For example, having spelling representations for iphone 6, iphone6, and iphone six, all with the same pronunciation, reduces the detection accuracy. The transcript can have any of these variants selected every time. Too many spelling representations impact the search and categorization where all the representations must be considered.
Validate and fine tune

Two weeks after the new language model is installed, review the transcribed interactions in Speech Analytics or RTSA. Search for interactions with the terms and phrases that you boosted or modified. You should see a significant improvement.

You will usually see an increase in the term or category trends from the date when the new model was installed. After the new model is deployed, Speech Analytics transcribes the terms that were boosted more accurately. New interactions without category hits will now show hits, as designed.

A quick way to validate that the changes are operational is to search for a new term that you added to the dictionary, and validate that the system transcribes the term accurately and that the search returns correct results.

You can also compare the results of a specific phrase before and after the change in the language model was installed, to ensure that the new model is successfully integrated.