In 2020, the number of digital voice assistants in use reached 4.2 billion. It’s predicted that in 2024 there will be more than 8.4 billion devices with voice assistants. Thanks to this tendency, the technology does not stand still and is rapidly developing. Therefore, new products are being released, offering a more sophisticated user experience.
Mobile users, in particular, are in favor of voice assistants. They’re three times more likely to use voice search, while 20% of mobile queries are done with speech search.
What is a new generation of mobile voice assistants going to offer to meet the increasing expectations and demands of consumers? Let’s look at one of the products that gets user experience to a new level.
It’s a personal mobile assistant that is able to answer calls and helps avoid missed ones. Moreover, it can protect a user’s device from fraud and spam. The technology developed by Neuro.net is unique as you don’t need to install the app — the voice assistant is offered by mobile service providers and is connected to their mobile core network.
Why the assistant is unique
Mobile assistant technology is based on machine learning. Its voice, dialogues and other features are almost indistinguishable in comparison with human characteristics. To achieve these similarities Neuro.net had to “teach” the voice assistant the following:
- Imitation of emotions. The assistant had to learn how to express natural human emotions in conversations;
- Background noise. The agent switches on office or street noise;
- Pauses in speech. It’s able to reflect human speech, that includes natural pauses;
- Other attributes of human speech. For example, intonation, crutch words, natural errors, speech flow and others.
The AI-agent can answer calls in several scenarios. For instance, when the user:
- is out of the coverage area;
- dismisses the call;
- switches the phone to airplane mode;
- turn the phone off.
The mobile assistant picks up the phone, introduces itself and asks a caller what he or she needs from the user. Then it checks if there is:
- spam
It detects ads and if they’re out of the user’s interests, the agent ends the call;
- important information
The assistant talks and questions a person, who wants to leave information for the user. The summary of the dialogue is stored and sent to the user’s messenger or mobile operator account;
- fraud
The agent is taught to identify scammers. It records the conversation and warns the user about scam calls.
How training of the assistant is conducted
Unlike most virtual agents with synthesized speech, the mobile assistant developed by Neuro.net has emotional and natural human speech. Here is how it’s trained to maintain human conversations:
- Initial training consists of creating correct responses for spam and fraud protection as well as for important messages for the user. The team provides the assistant with a knowledge base and then trains it to understand human speech. The assistant converts the speech during the conversation into data, which can be processed and compared to the knowledge base. After analyzing similar cases, the agent chooses the right response.
- After the first stage, the test version is released. The AI-agent listens to real dialogues and accumulates knowledge. If an unknown expression is found, the assistant compares it to its knowledge base. Depending on the percentage of matches, it uses the most suitable answers. Finally, the new expression is added to the knowledge base and then used in future interactions. These two stages take about three weeks.
- For skills improvement and additional training, the assistant is taught for one more week. It’s then launched for a small group of users.
During the test run around ten thousand calls are needed, so the agent is trained to understand human speech with 95% accuracy. When this number is reached, it’s able to train by itself. The final version is ready in about two months after the project launch.
The future of voice assistants
Here are the key predictions of voice assistants features in the nearest future:
- Synchronized contact lists
All contacts of the user can be added to the whitelist of the personal assistant. Thus, the agent will be able to use them during conversations.
- Synchronized calendar
The assistant will operate according to calendar notes. For instance, if a caller wants to adopt an appointment date, the agent will ask the user for permission. The user will only need to approve or reject the suggestion.
- Reconnecting to the conversation
If the connection is lost during a conversation, the assistant informs a person on the other end, asks not to hang up or calls back after it’s restored.
