Speech technologies is a part of everyday life. Voice assistants have been helping us surf the Internet, make purchases, and operate home appliances for many years. However, the market is developing and it constantly offers new products and solutions.
Now, let’s talk about a new product: a personal mobile assistant that can answer calls, helps avoid missed calls and protects against scammers and spam. Unlike usual personal assistant apps, the Voctiv technology does not require app installation as it is provided in partnership with wireless carriers and has a connection to their mobile core network.
Skills and uniqueness
This personal mobile assistant is developed based on Voctiv virtual agents’ technologies. Thanks to machine learning, it is difficult to distinguish the AI assistant from a human assistant. It can also conduct complex dialogues. To achieve this we use:
- emotion imitation
expressing emotions that are natural for human conversations; - background noise
superimposing office or street sounds; - pause adjustment
pauses in agent’s speech that mirror human speech; - unique features of speech
small errors, parasite words: details that characterize human speech.
The user sets up an assistant that can answer calls in several situations:
- when the user is out of range;
- when their phone is switched to airplane mode;
- when their phone is turned off;
- when the user dismisses the call.
The assistant introduces itself as a personal assistant and asks what information is needed by the user. After that the AI uses one of the following skills:
- spam protection
If the assistant identifies an ad that does not interest the user, it ends the call; - important information storage
If a call is from a person who wants to leave information for the user, the assistant starts a dialogue. It asks leading questions and sends a summary of the conversation to the user’s messenger. - fraud protection
The assistant can recognize scammers. In this case, it records the conversation and also warns the user.
Scenarios and Training
The Voctiv assistant speaks emotionally and in a human voice. This feature distinguishes it from solutions that use ordinary synthesized sound: such speech immediately reveals a virtual agent in a conversation.
But how is an assistant taught to respond correctly to human phrases? In the first stage, the implementation team creates a conversation flow for common skills. These skills represent correct responses:
- to spam;
- to messages that may be of interest to the user (for example, a call from a health care provider);
- to calls from friends or to actions of fraudsters.
In fact, the team creates the skills and knowledge base for the assistant and then trains the assistant to understand a natural language. During the conversation the system converts the speech into data suitable for processing and compares it with the knowledge base. Then it finds similar cases in the knowledge base and selects an appropriate response option.
After the initial training, the pilot version is launched. The assistant participates in real dialogues and accumulates knowledge. For example, if the assistant encounters a new phrase it compares it to its knowledge base. If it gets a high percentage of matches, the answer is used for similar cases. A new phrase is added to the knowledge base and then used in future conversations.
The development of the first stage takes about three weeks. Skill Improvements and additional training of the assistant take another week. After that, the assistant is launched for a small group of users. It takes about ten thousand calls during the pilot run to train the agent to understand human speech with an accuracy of close to 95%. After that, the agent is able to train independently. The final version of the assistant is ready in about two months after the launch of the project.
Plans and prospects
In addition to what voice assistants can do today, there are several skills that they will master in the near future. They are:
- Synchronization with a list of contacts
It is planned that all contacts of a user will be added to the white list so that the assistant will be able to use them in dialogs; - Synchronization with calendars
The assistant will be able to operate according to the calendar based on results of a conversation. For example, if someone calls to confirm or make an appointment, the assistant will prompt the user to add it to their planner. The user will only need to approve or reject the proposed changes; - Restoring the connection
The assistant can call back or reconnect to the conversation if the connection is lost. For example, if the user enters an underground passage the assistant tells the person on the other end that there is a problem with connection. Then it asks them not to hang up, as the conversation will continue soon.
