Push-to-Talk vs Always Listening for AI Devices

Push-to-talk vs always listening is really a question about who opens the microphone. With push-to-talk, you do, by holding a button. With a wake word, a small detector on the device listens for a phrase and opens it for you. An always-listening device processes sound all the time. If you are building an AI gadget with Meta's Muse Gadgets SDK today, the choice is simple: every voice board in the SDK uses push-to-talk.

The three listening modes in plain terms

All three modes can run on the same hardware: a microphone, a chip and a network connection. What changes is the rule for when sound gets recorded and sent anywhere.

Push-to-talk. The mic records only while you hold a button. On a Home Assistant Voice Preview Edition running the Muse firmware, holding the center button records a voice note that Muse answers in the Muse app.
Wake word. The device runs a small detector nonstop and starts recording after it hears a set phrase. Home Assistant's own firmware for the same Voice PE does this on the device with an engine called microWakeWord, listening for "Okay Nabu", "Hey Jarvis" or "Hey Mycroft".
Always listening. The device processes all the audio it hears, for example to transcribe a meeting or react to sounds. None of the Muse gadget builds work this way.

So the same $69 Voice PE can be a wake-word speaker for Home Assistant or a push-to-talk speaker for Muse. It depends on the firmware you flash.

Push-to-talk vs always listening vs wake word, side by side

What matters	Push-to-talk	Wake word	Always listening
What opens the mic	Your button press	A phrase the device detects	Nothing; it stays open
Privacy	Records only while you hold the button	Records after each trigger, including false ones	Depends on where the audio is processed
False triggers	Only accidental presses	A known problem on smart speakers	Not a trigger problem; everything is heard
Battery	Best; the board waits for a button	Mic and detector run all the time	Highest; audio is processed constantly
Hands-free	No	Yes	Yes
Muse gadgets today	Every voice board	Not in the SDK docs	Not in the SDK docs

The tradeoff is convenience. Push-to-talk wins on almost everything except the moment your hands are covered in flour.

Privacy: who controls the microphone

Push-to-talk gives you the simplest rule to explain to everyone in the house: nothing is recorded unless someone is holding the button. On Muse gadgets, that voice note goes to Muse, which transcribes it and answers in writing. Muse's homepage says your conversations "are not shared with Meta's ad systems".

A wake-word device has a harder job. Its detector has to hear everything to find the phrase. On the Voice PE's stock firmware, Home Assistant says that detection runs on the device. The weak spot is the mistakes: every false trigger starts a recording of whatever is said next.

Hardware switches help with either mode. Home Assistant says the Voice PE's mute switch "physically cuts power to the microphones". In the Muse build, turning the switch on also gives the center button back its setup role, so you can confirm pairing or hold it for 5 seconds to reset.

False triggers: how often wake words misfire

Wake words misfire because TV, music and conversation are full of sounds that resemble the phrase. Researchers from Northeastern University and Imperial College London tested this by playing 125 hours of Netflix shows to smart speakers including the Echo Dot, Google Home Mini, HomePod and Harman Kardon Invoke. They measured an average of 1.5 to 19 accidental activations per device per day, depending on the device, and the longest ones lasted 20 to 43 seconds.

Push-to-talk avoids this almost entirely. Its version of a false trigger is an accidental press, such as a pocket device bumping against your keys. That is easier to notice and easier to prevent, for example by carrying the device in a case.

If you add your own wake word to a gadget later, plan for misfires from day one: give people a visible sign that the mic is open and a simple way to mute it.

Battery: why push-to-talk lasts longer

A wake word needs the microphone and some processing running every second the device is on. Push-to-talk lets the board wait for a button. That matters most on pocket boards, which have small batteries: the M5Stack StickS3 has a 250 mAh cell, and the AIPI Lite comes with a 450 mAh battery module.

The Muse firmware helps here too. When you power the StickS3 off from Muse, it turns off the screen, speaker and audio codec and puts the ESP32-S3 into deep sleep, and either button wakes it. M5Stack also advises keeping the speaker volume below 75% on battery alone "to avoid unexpected device reboot".

Plugged-in boards care less. The Voice PE runs from USB-C at 5 V and 2 A, so for that board the choice is purely about privacy and convenience.

Which Muse boards do which

Here is how you talk to each voice board hackshop supports, and where the reply shows up. Replies from Muse arrive as text, and boards with screens show them as captions. For spoken answers, you send each reply to a text-to-speech API of your choice. On boards with PSRAM, the SDK's start_tts function is where you hook it in.

Board	How you talk	Where replies show	Price checked
Home Assistant Voice PE	Hold the center button	In the Muse app; status on the LED ring	$69 MSRP
M5Stack StickS3	Front button	Captions on a 1.14" screen	$21.50
AIPI Lite	Push-to-talk plus a two-button menu	Captions on a 128×128 screen	$35.99
Waveshare ESP32-S3-Touch-AMOLED-1.75C	Push-to-talk button	Captions on a round 1.75" AMOLED	$39.99 to $41.99
Seeed SenseCAP Watcher	Press the wheel	Captions on a round 1.45" screen	$60.99
Waveshare ESP32-C6-Touch-AMOLED-1.8	BOOT button	Text that scrolls past	See board page
M5Stack StickC Plus2	Front button	Captions; buzzer only	End of life

The status boards, such as the ESP32-C5 DevKitC-1, the ideaspark 1.9" screen, the SenseCAP Indicator and the reTerminal e-paper displays, have no push-to-talk in the Muse build. The Raspberry Pi gadgets run commands rather than voice. None of the boards lists a wake word.

How to choose a listening mode for your build

Shared rooms and kitchens: push-to-talk with a hardware mute is the easiest setup to trust. The Voice PE gives you both.
On the go: a pocket board with a front button works like a walkie-talkie for your agent. The StickS3 is a low-cost full-UI option at $21.50.
At your desk: pick a board with a screen so you can read replies, such as the Waveshare 1.75C or the Watcher. The desk companion comparison covers the three main picks.
You truly need hands-free: be honest about the work. A wake word is not in the Muse SDK today, so you would be adding and tuning it yourself.

Still unsure? The Muse board picker quiz suggests a board from a few questions, and the Muse board guide compares every supported board side by side.

Questions people ask

Do Muse gadgets support a wake word?

Not according to the SDK documentation. Every board in the Muse ESP32 SDK that has a microphone is listed with push-to-talk, and none is listed with a wake word. The code is open source under Apache 2.0, so you could add one, but you would be building and tuning it yourself.

Is push-to-talk more private than a wake word?

In most homes, yes. Push-to-talk records only while someone holds the button, so the TV or a conversation can't set it off. A wake-word device can also be private, but every misfire starts a recording of whatever is said next.

Will my gadget answer out loud?

Not by default. Muse replies with text, which boards with screens show as captions and the Voice PE shows in the Muse app. For spoken answers, you send each reply to a text-to-speech API of your choice and play the audio it returns.

Does the Voice PE mute switch still work with Muse firmware?

Home Assistant says the switch physically cuts power to the microphones, which is a hardware action rather than a software setting. In the Muse build, turning it on also returns the center button to its setup role for pairing and resets.

Sources

Prices and details checked October 5, 2026.