Pre-built Language Operators
Warning
Public Beta
Voice Intelligence is currently available as a public beta release. Some features are not yet implemented and others may be changed before the product is declared as Generally Available. Beta products are not covered by a Twilio SLA.
Learn more about beta product support.
Language Operators turn transcripts into structured information using a variety of techniques, including machine learning. Pre-built Language Operators (PBLOs) have been created by Twilio or use third-party AI models. Twilio made Operators are trained across a wide swath of data and typically map to data types that are agnostic to use-case or industry. PBLOs can't be modified or made more specific.
The following Pre-built Language Operators are currently available for use.
| Operator Name | AI model | Operator Action | Description |
|---|---|---|---|
| Sentiment Analysis | OpenAI GPT-3.5 | Classify | Classifies the sentiment of the call as Positive, Negative or Neutral |
| Summarization | OpenAI GPT-3.5 | Text generation | Creates a summary of the conversation |
| Entity Recognition | Amazon Comprehend | Classify-Extract | Extract entities from the conversation |
| Agent Introduction | Twilio | Phrase matching | Checks if a participant introduced themselves during the conversation |
| Escalation Request | Twilio | Phrase matching | Checks if a participant requested an escalation during a conversation |
| Outbound Call Disposition | Twilio | Classify | Classifies the outcome of an outbound call into a general category |
| Recording Disclosure | Twilio | Phrase matching | Checks if a participant disclosed that the call was being made on a recorded line |
| Voicemail Detection | Twilio | Classify | Classifies whether a call went to voicemail or was picked up by a human |
| Unavailable Party Detector | Twilio | Classify | Classifies if the party being called is unavailable |
| Do Not Contact Me | Twilio | Phrase matching | Checks if a participant requested to avoid being contacted |
| Non English Call | Twilio | Classify | Classifies whether or not the conversation is in English |
| Password Reset | Twilio | Phrase matching | Checks if a participant requested a password reset |
| Call Transfer | Twilio | Classify | Classifies if the call was transferred to another agent |
Info
To find out more about Language Operators and the actions they can perform, check out the key concepts page.
Action: Classify
Base Model: OpenAI GPT-3.5
Info
The first 12,000-13,000 words of the transcript are used to determine the sentiment of the conversation.
The Sentiment Analysis operator determines the sentiment of the conversation. The Sentiment Analysis operator doesn't analyze a specific participant of the conversation.
| Class | Description |
|---|---|
positive | The sentiment of the call was positive |
negative | The sentiment of the call was negative |
neutral | The sentiment of the call was neutral |
mixed | There was more than one sentiment during the call |
Action: Text generation
Base Model: OpenAI GPT-3.5
Info
The first 12,000-13,000 words of the transcript are used to generate the summary of the conversation.
The Summarization operator generates a summary with the key points and important information discussed in the conversation.
Action: Classify-Extract
Base Model: Amazon Comprehend
The Entity Recognition operator recognizes unique names such as locations, organizations, consumer goods or people, quantities and times.
| Entity | Description |
|---|---|
Person | Individuals, groups of people, nicknames, fictional characters |
Location | A specific location, such as a country, city, lake, building, etc. |
Organization | Large organizations, such as companies, government, sports teams, etc. |
Consumer_good | A branded product |
Date | A full date (e.g. 10/20/1997), day (Wednesday), month (September), or time (10:15 a.m.) |
Quantity | A quantified amount, such as currency, percentages, numbers, bytes, etc. |
Action: Phrase matching
The Agent Introduction operator determines if an agent introduced themselves on a call.
The Agent Introduction operator doesn't detect a specific type of introduction. If you want to determine whether or not an agent introduced themselves in a specific way, we recommend that you create a Literal Spot operator that looks for the introductions that you are attempting to find.
| Class | Description |
|---|---|
true | The operator spotted an introduction |
false | The operator did not spot an introduction |
Action: Phrase matching
The Escalation Request operator determines whether or not a customer requested an escalation during a call. It captures events like a customer asking to speak to a manager or a supervisor.
| Class | Description |
|---|---|
true | The operator spotted an escalation request |
false | The operator did not spot an escalation request |
Action: Classify
The Outbound Call Disposition operator determines the outcome of an outbound call. It is is geared toward customers who are running outbound dialing campaigns and have a set of outcomes that are focused on campaign list management.
| Class | Description |
|---|---|
GhostCall | The call was picked up and one participant does not provide any audio |
IVR | The call was picked up by an automated Interactive Voice Response (IVR) system |
NoContent | There was no audio during the call |
Ordinary | The call was picked up by a human and proceeded normally |
StopContactingMe | The call was picked up by a human who requested to be removed from a campaign list or requested to be placed on a do-not-call list |
Voicemail | The call was picked up by a voicemail system |
WrongNumber | The call was picked up by a human who indicated that it was a wrong number |
Action: Phrase matching
The Recording Disclosure operator determines whether or not an agent notified the customer that the call was being recorded. Typically, this operator is used on outbound calls where it doesn't make for a good customer experience if you use a recording to announce that the call is being recorded before a human agent greets the customer.
The Recording Disclosure operator doesn't match a specific type of recording disclosure. If you have a specific disclosure that you want to ensure that agents use, we recommend creating a Literal Spot operator to match the required phrasing.
| Class | Description |
|---|---|
true | The operator spotted a participant mentioning that the call was being recorded |
false | The operator did not spot a participant mentioning that the call was being recorded |
Action: Classify
The Voicemail Detection operator determines whether or not a call went to voicemail. It is geared towards customers who are running outbound dialing campaigns and want to use the results to determine whether to call a specific customer back, and to assess the best time to make that call.
Unlike the Outbound Call Disposition operator, the Voicemail Detection operator only indicates whether a call was handed to a voicemail system. It does not use acoustic features to determine whether or not the call was actually picked up by the voicemail system.
| Class | Description |
|---|---|
voicemail | The call was picked up by a voicemail system |
not_voicemail | The call was not picked up by a voicemail system |
Action: Phrase matching
The Do Not Contact Me operator determines whether or not a customer requested to stop being contacted during a call.
| Class | Description |
|---|---|
true | Customer requested to stop contacting this number |
false | Customer did not request to stop contacting this number |
Action: Classify
The Non English Call operator determines whether or not the conversation was in English.
| Class | Description |
|---|---|
NonEnglishCall | Participants spoke in a non-English language |
EnglishCall | Participants spoke in English |
Action: Classify
The Password Reset operator determines whether or not a call participant requested to change their password.
| Class | Description |
|---|---|
NoPasswordReset | Password Reset was not requested |
PasswordReset | Password Reset was requested |
Action: Classify
The Call Transfer operator determines whether or not the call was transferred.
| Class | Description |
|---|---|
NoTransfer | The call was not transferred |
Transfer | The call was transferred |
Below are the current AI Nutrition Facts labels for Voice Intelligence. Voice Intelligence includes three Base Models (OpenAI-GPT-3.5, Amazon Comprehend, and Twilio's own AI model).
For OpenAI-GPT-3.5 and Amazon Comprehend, Customer Data is not used to train the Base Models. For Twilio's own AI model, if you opt in to accept the Data Logging Terms, then Twilio may use your Customer Data to train its own Base Model. For more information, please refer to the Twilio Voice Intelligence: Data Logging Consent Addendum.
To learn more about Twilio's AI Nutrition Facts and how to read these labels, please visit https://nutrition-facts.ai/.
