Speech-to-text (ASR) models are AI models that turn spoken words into written text. ASR stands for automatic speech recognition. In many products the model does not stop at the text. It also works out what the speaker wants, so that the product can answer. Such models run in car voice assistants, restaurant ordering systems and household devices, and they are now combined with conversational AI models that reply in natural language.
The model receives sound as a stream of numbers. A microphone measures the air pressure 16,000 times a second. The model cuts this stream into short frames, about 10 to 25 milliseconds each, and describes each frame with numbers for its pitch and loudness at different frequencies. It then predicts the most likely words, using what it learned from thousands of hours of recorded and transcribed speech. A good model also uses context. So it prefers 'petrol pump' over a similar-sounding phrase that makes no sense in a car.
Next comes understanding. The text is passed to a model that finds the intent, meaning what the speaker wants, and the details, called slots. For 'find the nearest petrol pump', the intent is 'find a place' and the slot is 'petrol pump'.
Take, for example, a driver on the Mumbai–Pune Expressway in a car with an in-car voice assistant. The fuel light comes on. She presses the button on the steering wheel and says, in a mix of Hindi and English, 'Sabse paas wala petrol pump dikhao.' The microphone records 2 seconds of speech, which is 32,000 measurements. The model returns the text in about 300 milliseconds after she finishes. The understanding step finds the intent 'find a place' and the slot 'petrol pump'. The car's map finds a pump 4 km ahead at the next exit, and the assistant announces it aloud. Her hands stay on the wheel.
Cars are hard for speech models. The engine, the wind, the air conditioner and the passengers all add noise. Indian drivers also mix languages in one sentence. So carmakers test models with recordings made inside moving cars, and the models are trained on such recordings. The same technology takes food orders at restaurant counters and through drive-thru headsets.
The general idea is explained on the page on speech recognition. This page is about the model itself, tuned for noisy places and quick answers. It is part of speech and voice AI, beside real-time voice AI, which produces the spoken reply.