You ask Alexa for the weather, tell Siri to set a timer, or ask Google to add milk to the shopping list, and a second later it answers. It feels like talking to someone. Underneath, it’s a chain of separate steps, some on the device in front of you and some on servers far away.
Step 1: The wake word
A smart speaker or phone is always listening, but only for one thing: its wake word, such as “Alexa”, “Hey Siri” or “Hey Google”.
A small, low-power chip runs a tiny model that compares incoming sound with the wake word, over and over. It keeps only a few seconds of audio in a rolling buffer and throws it away unless it hears a match. When it does, the device lights up and starts recording your request.
False triggers happen, which is why assistants sometimes wake up to a word on TV. Most let you review and delete recordings in their app.
Step 2: Turning speech into text
Your request, such as “what’s the weather in New York tomorrow?”, is now a short audio clip. Speech recognition (also called speech-to-text) turns it into written words.
For years this happened almost entirely in the cloud: the device compressed the clip and sent it to the company’s servers. Phones now do much of it on the device itself, which is faster and keeps more audio off the internet. Smart speakers still send most requests to the cloud.
Step 3: Working out what you want
Next the assistant has to understand the text. The classic approach, used by assistants for most of the 2010s, breaks the sentence into an intent and details:
- Intent: get weather forecast
- Details: place = New York, day = tomorrow
Each intent was linked to a fixed action, here a call to a weather service. That’s why older assistants were great at timers and music but fell over when you phrased something in an unexpected way: if a request didn’t match an intent someone had built, they couldn’t help.
Step 4: The answer, spoken back
The assistant writes its reply as text, and text-to-speech turns it into a voice. Modern voices are generated by neural networks trained on many hours of recorded speech, which is why they no longer sound robotic. The audio is played on your speaker, usually within a second or two of you finishing your sentence.
What large language models changed
Since 2025, all three big assistants have been rebuilt around large language models (LLMs), the same kind of AI behind chatbots like ChatGPT and Gemini:
- Amazon launched Alexa+, an LLM-powered version of Alexa, on compatible Echo devices.
- Google began replacing Google Assistant with Gemini on phones, speakers and smart displays.
- Apple announced a partnership with Google in January 2026 to use Gemini models in the next generation of its own AI models, which power the rebuilt Siri.
The difference is in step 3. Instead of matching your words to a fixed list of intents, an LLM can follow a loose, multi-part request (“find a dinner recipe that uses the chicken in my fridge and add the missing ingredients to my list”), remember what you said a moment ago, and answer open questions in full sentences.
It isn’t a clean upgrade. LLMs are slower and more expensive to run than the old systems, and they can state wrong things confidently. Reviewers in 2026 found the new assistants better at conversation but sometimes worse at the simple, reliable tasks people use most, like timers and smart home control. Google, for one, keeps simpler, faster models for basic commands and calls on the bigger model for open-ended chat.
What this means for privacy
More capable assistants generally need more processing in the cloud. In 2025 Amazon removed the option for some Echo devices to process requests locally, so recordings now go to its servers. Whichever assistant you use, it’s worth opening its app once to check:
- whether recordings are saved, and for how long
- whether people at the company can review recordings to improve the service
- how to delete your history
All three let you turn off recording storage or delete it automatically after a set time.
