How can we help?

AI Trainer: FAQ

Follow

 

Index

 

What does AI Trainer do?

How can AI Trainer be installed?

How can I configure access to AI Trainer?

Why doesn’t my account show any models to start with?

Which roles does AI Trainer include per default, and how are they configured?

Speech to text model

What AI products were integrated with AI Trainer? 

What type of models are supported?

What languages are supported in AI Trainer models?

Which model providers are supported?

Why can’t I test transcriptions if I have an STT model configured?

Where can I click to train the custom vocabulary?

After adding a new phrase or editing an existent one, a message displays indicating there is a validation error. What should I do?

What are utterances, intents, and entities?

How can I know which intents I should add?

How to ramp up the model and improve existing intents? What is the agreement process?

 

 

 

 

AI Trainer: Overview

What does AI Trainer do?

image3.png

Talkdesk AI Trainer™ is the first human-in-the-loop tool to be included in the operational flow of contact centers, allowing for continuous improvement of AI models. With AI Trainer, non-technical staff can easily apply their domain expertise through clicks, rather than code.

Crowdsourcing these improvements next to a larger audience (for example, your agents) allows you to gather more feedback and therefore increase the performance of your models at a faster pace and at a lower cost. 

Through our quality system, you have the final say regarding the choice of improvements to be introduced on your models and the products that make use of those models.

AI Trainer enables you to successfully resolve more cases through automation and improve accuracy, increasing customer satisfaction and decreasing the cost per case.

 

AI Trainer: Installation and configuration

How can AI Trainer be installed?

AI Trainer is available to all customers. The application is preinstalled by default in your account. After installation and access configuration (adding users or user groups to allow-list during installation), AI Trainer can be accessed from My Apps or via direct link. 

If you notice that AI Trainer is not installed in your account, please contact the Administrator of your account, so access can be granted.

If you are an Administrator experiencing this issue, please contact your Custom Success Manager or Talkdesk Support

 

How can I configure access to AI Trainer?

image8.png

During installation, it is possible to configure which users, by name, or which user groups can access AI Trainer. If at any other given moment it is necessary to configure access, this can be done via Settings in the AI Trainer My Apps app (gear icon in top-right corner of application icon under My Apps). 

 

image1.png

 

Why doesn’t my account show any models to start with?

If there are no model cards shown on the Models' page, it means no models or integrations are configured for the account. Please contact your Custom Success Manager or Talkdesk Support to create and configure a new model (via AI Settings).

 

Which roles does AI Trainer include per default, and how are they configured?

By default, when a Talkdesk account is created, the following roles are created: Administrator, Supervisor, and Agent. For more details, please click here

 

image7.png

 

It is also possible to add, to a custom role, specific AI Trainer permissions. This is possible for Admins, by accessing Talkdesk Classsic (Admin section > Roles and Permissions tab). Select Edit permissions on the custom role and expand the details of AI Trainer to select the permissions you wish to allow.

Below, you can find a description of the different permissions and the default permissions that each role has on AI Trainer. Please note that, currently, Administrators and Supervisors have the same permissions. 

 

Models 

  • View: read-only access to the Models page.

Intent models

  • Trends: 
    • View: read-only access to the Trends page.
  • Inbox:
    • View: read-only access to the Inbox page.
    • Label data: classify utterances with their corresponding intents.
  • Intents: 
    • View: read-only access to the Intents page.
    • Create: create new intents.
    • Edit: edit intents.
    • Delete: delete intents.
  • Entities: 
    • View: read-only access to the Entities page.
    • Create: create new entities.
    • Edit: edit entities.
    • Delete: delete entities.
  • Quality check: recommended to give access only to advanced users
    • View: read-only access to the Quality check page.
    • Review intent: approve or decline the classification of intents. 
  • Launch: recommended to give access only to advanced users
    • View: read-only access to the Quality check page.
    • Launch intent: deploy the changes made to intents to the model and to the products that make use of the model.

Speech to text model

  • Vocabulary:
    • View: read-only access to the Vocabulary page.
    • Create: create new phrases.
    • Edit: edit phrases.
    • Delete: delete phrases.
    • Test: test vocabulary.

Clustering (this feature is in Early Access)

  • View: read-only access to the Clustering page.

AI Trainer | Roles and Permissions

Areas

Roles

Administrator / Supervisor

Agent

Models

View

       

View

 

Intent Model

Trends

View

       

View

 
 

Inbox

View

Label data

     

View

Label data

 

Intents

View

Create

Edit

Delete

 

View

 
 

Entities

View

Create

Edit

Delete

 

View

 
 

Quality check

View

Review intent

         
 

Launch

View

Launch intent

         

Speech to text Model

Vocabulary

View

Test

Create

Edit

Delete

View

 

Clustering Model

 

View

       

View

 

 

What AI products were integrated with AI Trainer?

AI Trainer has been used and tested in integration with Talkdesk Virtual Agent, Agent Assist, Interaction Analytics, Knowledge Management,  and Quality Management Assist.

 

What type of models are supported?

At the moment of General Availability (GA) release, AI Trainer supports Speech-to-Text models and Intent detection models (with entity extraction as well). 

Clustering models are available only via Early Access program.

 

What languages are supported in AI Trainer models?

Following are the supported languages.

Note: It might vary considering the chosen provider.

  • Streaming mode:
    • Mandarin Chinese – Mainland (zh-CN).
    • Australian English (en-AU).
    • British English (en-GB).
    • US English (en-US).
    • French (fr-FR).
    • Canadian French (fr-CA).
    • German (de-DE).
    • Italian (it-IT).
    • Japanese (ja-JP).
    • Korean (ko-KR).
    • Brazilian Portuguese (pt-BR).
    • US Spanish (es-US).
  • Batch transcriptions: if your STT model is configured for batch transcriptions, live testing of the STT model with your own voice (Test vocabulary) won’t work. This feature is only supported in streaming mode. 
    • Gulf Arabic (ar-AE).
    • Modern Standard Arabic (ar-SA).
    • Mandarin Chinese – Mainland (zh-CN).
    • Dutch (nl-NL).
    • Australian English (en-AU).
    • British English (en-GB).
    • Indian English (en-IN).
    • Irish English (en-IE).
    • Scottish English (en-AB).
    • US English (en-US).
    • Welsh English (en-WL).
    • Spanish (es-ES).
    • US Spanish (es-US).
    • French (fr-FR).
    • Canadian French (fr-CA).
    • Farsi Persian (fa-IR).
    • German (de-DE).
    • Swiss German (de-CH).
    • Hebrew (he-IL).
    • Indian Hindi (hi-IN).
    • Indonesian (id-ID).
    • Italian (it-IT).
    • Japanese (ja-JP).
    • Korean (ko-KR).
    • Malay (ms-MY).
    • Portuguese (pt-PT).
    • Brazilian Portuguese (pt-BR).
    • Russian (ru-RU).
    • Tamil (ta-IN).
    • Telugu (te-IN).
    • Turkish (tr-TR).

Training intent detection models:

Note: Please contact your Custom Success Manager or Talkdesk Support to know the local variants as well as other supported languages which are available in Preview mode.

  • Bengali (bn).
  • Chinese - Cantonese (zh-HK).
  • Chinese - Simplified (zh-CN).
  • Chinese - Traditional (zh-TW).
  • Danish (da).
  • Dutch (nl).
  • English (en).
  • French (fr).
  • German (de).
  • Hindi (hi).
  • Indonesian (id).
  • Italian (it).
  • Japanese (ja).
  • Korean (ko).
  • Norwegian (no).
  • Polish (pl).
  • Portuguese - Portugal (pt).
  • Russian (ru).
  • Spanish (es).
  • Swedish (sv).
  • Thai (th).
  • Turkish (tr).
  • Ukrainian (uk).

 

Which model providers are supported?

AI Trainer supports AWS Transcribe for STT models. For intent models, AI Trainer supports Google Dialogflow (both the ES and the CX editions), Talkdesk’s NLU services, and, soon, Amazon LEX.

 

AI Trainer: Speech-to-Text Models

Why can’t I test transcriptions if I have an STT model configured?

Before testing the transcription with custom vocabulary, firstly, you need to add at least one custom phrase. Even with added phrases, if your STT model is configured for batch transcriptions, live testing with your own voice won’t work as this feature is supported only in streaming mode. 

 

image6.jpg

Note also that while the STT model is training (after a new phrase is added or a phrase is edited) the option to test transcriptions becomes temporarily unavailable.

 

Where can I click to train the custom vocabulary?

Custom vocabulary is trained automatically after each insert of a new phrase.

 

After adding a new phrase or editing an existent one, a message displays indicating there is a validation error. What should I do?

Start by identifying which phrase contains the error.

 

image10.png

The number of the line that contains the error can be found on the error message [1]. In the example here shown, the error is on line 15.

Go to the corresponding page of the error [2]. Each page contains ten phrases. For example, if the error is on line 15, you need to go to page two to find it.

Please note that the title of the table (where the phrases are listed) counts as line number one [3]. Considering this and the particular example here shown, the line where the first phrase is listed on page 2 is line number 11. Now it is simple to identify line 15, where the error (two hyphens in a sequence in “Sounds like”) is.

 

image9.png

Most errors are space characters or some characters outside the allowed character set. Follow the validation error and correct the phrase. Common mistakes include:

  • Space between words in “Phrase” or “Sounds like”.
  • Two hyphens in a sequence in “Phrase” or “Sounds like”.
  • Using numbers as digits in “Sounds like”.

 

AI Trainer: Intent Models

What are utterances, intents, and entities?

In AI Trainer, every sentence that is being spoken during a call, either by the caller or by the agent, and then transcribed, is an utterance. One synonym for utterance that is used throughout the application is simply a phrase.

The intended action that the speaker has expressed in the utterance is an intent. 

Entities are metadata associated with the intent. If with the desired action a speaker mentions specific products, places, qualities, among others, that information can be extracted in the context of the intent detected.

 

How can I know which intents I should add?

The recommendation is that you set up, at least, three to five common intents relevant for your business and that represent also the frequent work of your contact center, like Cancelation, Escalation, Refund, Inquiry, among others. You can refine them later into more granular topics. 

For each intent created, the recommendation is to add 10-15 training phrases to start with, so that this intent can be correctly identified by the model. 

For the training phrases, the recommendation is to use different phrase formulations and keywords, so that they represent what customers and/or agents most often mention in their interactions within the specific intent they belong to.

 

How to ramp up the model and improve existing intents? What is the agreement process?

 

image4.png

AI Trainer ensures that agents, as domain experts, label data. To avoid changes being deployed to production accidentally (which would affect the model’s performance and therefore the product(s) that makes use of this intent model), there is an Agreement process in place. 

This process is a quality system that makes sure that, while agents can make suggestions for a change of intent classification, only a Supervisor or Administrator can review them (approve or decline). This is done on the Quality Check page.

 

image2.png

Once this revision is done, the intent can then be deployed to production via the Launch page. It is possible to launch immediately the latest changes made to the intent, or to schedule when these changes will affect the model and the product(s) that are supported by this intent model.

All Articles ""
Please sign in to submit a request.