Glossary
What is a voice assistant?
In short
A voice assistant is a program you talk to: it recognizes speech, understands the request and answers in a synthetic voice.
AI agents · voice assistant
MMOOZZGG.com editorsChecked: 1 min read
For a child
Sometimes your hands are busy with clay, but you need to know the time. You just ask out loud, and the speaker answers. Inside it works a helper that hears the words, understands what you want and talks back. It is not a person, even if the voice sounds like one. So agree with the grown-ups about what you can talk to it about.
In plain words
A talk with a voice helper goes through several stages. First, the program turns your speech into text. Then it understands what you want: the weather, an alarm, a doctor’s appointment. Then it performs the action or finds the answer and says it aloud. Earlier, such helpers understood only preset commands. Now a language model sits in the middle, and you can talk to them almost as to a person: interrupt, ask again, change the topic. The same principle works in phone robots that call clients or take calls.
For business
In a company, a voice assistant most often works on the phone: it takes incoming calls, books appointments, confirms visits and answers frequent questions. It does not get tired at peak hours and does not go on vacation. The price is made of speech recognition and synthesis, model fees, telephony and integration with the CRM or booking system. The main limit is delay: if the pause is too long, the caller starts to interrupt or hangs up. There are legal questions too: call recordings contain personal data, and if the voice is used to identify a client, that is already biometrics. Before launch, test how the robot handles accents and noise on real call recordings.
For a specialist
A voice assistant is built either as a cascade or as a single speech model. The cascade includes voice activity detection, automatic speech recognition (ASR), a text dialog on a language model and text-to-speech (TTS). An example of an open recognition model is Whisper (Radford et al., 2022), trained on 680,000 hours of audio with multilingual and multitask labels. The weak points of the cascade are the accumulated delay and the loss of intonation at the text stage. The alternative is speech-to-speech models, such as Moshi (Défossez et al., 2024), where the user’s and the system’s speech are modeled as parallel streams; the authors report a theoretical delay of 160 ms and about 200 ms in practice. Engineering tasks: interrupting the answer (barge-in), detecting the end of a turn, noise reduction and work in the narrow band of a phone channel.
Example
A dental clinic chain connects a voice helper to its incoming line. In the evening it takes a call, finds a free slot with the right doctor, books the patient and sends a confirmation to a messenger.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map