ESP32-S3 and PSRAM

On-device wake words (such as microWakeWord in ESPHome or ESP-SR’s WakeNet) need the ESP32-S3’s AI instructions and several megabytes of PSRAM for audio buffers. 8 MB PSRAM is the comfortable target.

Microphones

One mic works at arm’s length. Two mics allow basic noise reduction, and a four-mic array with a voice processor gives true far-field pickup with echo cancellation across a room.

Codec and Speaker

An audio codec (such as the ES8311) handles both mic input and speaker output with good quality. A 1 W speaker is fine for spoken answers; for music, use a 3 W amplifier or a line output to a real speaker.

Software Support

For Home Assistant, check for ESPHome voice assistant support. For AI chatbots, look for boards with ready firmware and documentation. For music, ESP-ADF (Espressif’s Audio Development Framework) and Arduino audio libraries are the main options.

Feedback and Extras

An LED ring or screen shows when the device is listening, a mute button protects privacy, and a TF card slot stores sounds and music locally.

When a single microphone can’t hear you over the TV or from across the kitchen, you need a real far-field array. Seeed’s reSpeaker XVF3800 pairs a circular four-microphone array with the XMOS XVF3800 voice processor, which performs echo cancellation, beamforming, de-reverberation, direction-of-arrival, voice activity detection, AGC and noise suppression before the audio reaches the XIAO ESP32S3. Echo cancellation in particular lets you interrupt the device while it’s playing music. This is the bare-board version for custom enclosures, robots and kiosks. It has about 4.2 stars from 6 ratings. Seeed also sells a USB-only version (B0GVJ5YQ58) without the ESP32 for use with a PC or Raspberry Pi.

The reSpeaker Flex uses the same XVF3800 voice processor and XIAO ESP32S3 as pick 3, but in Seeed’s Flex format, which is designed to fit a wider range of product shapes and mounting positions. It’s the version to look at if you’re designing your own smart-speaker housing, a ceiling-mounted voice unit or a voice-controlled appliance and want far-field audio without designing a mic array yourself. It’s very new, with 5 stars from 3 ratings, so check Seeed’s wiki for the mechanical drawings before committing to a design.

The original Atom Voice measures just 24 × 24 × 17 mm and combines an ESP32 with a built-in microphone and speaker. It has the largest review base of the small voice devices here (69 ratings), though the 3.6-star average reflects its limits: the older ESP32 and tiny speaker make it best for short spoken replies with server-side wake-word detection, not music. Its Bluetooth also lets it work as a small wireless speaker. Choose it if the S3R version is out of stock or you want the cheapest possible satellite; otherwise the Atom VoiceS3R is the better buy.

If you want to talk to an AI chatbot on your desk without writing firmware first, LAFVIN’s AI Chatbot Kit ships with open-source chatbot firmware preloaded on an ESP32-S3, along with the microphone, speaker and display parts and step-by-step tutorials. It’s the friendliest entry point here for students and beginners, and you can reflash it later with your own ESP-IDF or Arduino code. It has about 4.2 stars from 39 ratings. Online AI services may require an account and an internet connection, so check what the preloaded firmware uses before you start. It also appears in our ESP32 starter kits guide.

M5Stack’s CoreS3 is a polished, enclosed ESP32-S3 controller with a touch screen, an integrated camera, dual microphones and a speaker, which makes it a ready-made platform for voice assistants that also show responses on screen or recognize faces. It’s a popular base for smart-display style Home Assistant voice projects and AI assistants with a visual face. It has about 4.8 stars from 14 ratings. It’s pricier than a bare board, but you get a finished case, battery bottom and M5Stack’s large library of examples.

RASPIAUDIO’s ESPMUSE Proto is an ESP32 development board with a speaker and microphone already on the PCB, and the listing highlights Home Assistant compatibility. It suits ESPHome media-player and voice projects, internet radios and talking notifications where you want a bigger speaker than the tiny Atom devices without wiring separate modules. It has about 3.8 stars from 56 ratings, with only 3 in stock when we checked. Check RASPIAUDIO’s documentation for the current ESPHome configuration before you buy.

The ESP32-Audio-Kit is built around the ESP32-A1S, an ESP32 module with an integrated audio codec, and it’s the classic board for Bluetooth speakers, internet radios and Wi-Fi music players using ESP-ADF or the Arduino audio-tools library. The board breaks out headphone and speaker outputs, onboard microphones and buttons for play, volume and mode (check the listing photos for your board revision; this is V2.2). It needs a solid 5 V supply of more than 1 A. It has about 3.9 stars from 42 ratings, with 4 in stock. Note that A1S boards have shipped with different codec chips over time, so confirm which one yours has before choosing a library configuration.

The INMP441 is the standard I2S microphone for ESP32 voice projects, and this 5-pack (about 4.5 stars from 94 ratings) gives you enough for several satellites. Because it outputs digital I2S audio directly, there’s no analog noise or ADC calibration to worry about: connect SCK, WS and SD to three ESP32 GPIOs, power it from 3.3 V, and tie the L/R pin to choose the channel. Two INMP441s on one I2S bus (one set left, one right) give you a stereo pair for basic noise handling. It works with ESPHome’s i2s_audio microphone component and ESP-IDF’s I2S driver.

The MAX98357A is the matching output half of a DIY build: an I2S amplifier with a built-in DAC that drives a small 4 or 8 Ω speaker at up to 3 W directly from the ESP32’s digital audio, with no separate DAC needed. This 2-pack has about 4.5 stars from 87 ratings. It can share the BCLK and LRCLK lines with an INMP441 or use a second I2S port, and it’s supported by ESPHome’s i2s_audio speaker and media-player components. Power it from 5 V for full volume, and keep speaker wires short to reduce noise.

If you don’t want to source a speaker separately, NULLLAB’s kit pairs an NS4168 I2S amplifier with a matching 3 W speaker, two sets per pack. The NS4168 does the same job as the MAX98357A (digital I2S in, amplified speaker out), so it uses the same three ESP32 audio pins and works with the same ESP-IDF and ESPHome I2S drivers. It has excellent reviews, about 4.7 stars from 41 ratings, and it’s a tidy way to finish a voice satellite or talking sensor quickly.

All-round voice and AI projects?

The Waveshare ESP32-S3 AI Smart Speaker board.

Home Assistant satellite in every room?

The M5Stack Atom VoiceS3R, or the original Atom Voice on a budget.

Need to be heard across the room?

A reSpeaker XVF3800 array with XIAO ESP32S3.

Voice with a screen?

The M5Stack CoreS3, or the LAFVIN AI Chatbot Kit for beginners.

Music player or Bluetooth speaker?

The ESP32-Audio-Kit or ESPMUSE Proto.

Building your own?

An ESP32-S3 board + INMP441 + MAX98357A or the NS4168 speaker kit.

Can an ESP32 do speech recognition on its own?

An ESP32-S3 can detect a wake word and recognize a small set of offline commands using Espressif’s ESP-SR library or ESPHome’s microWakeWord. Full speech-to-text and natural-language answers run on a server, such as Home Assistant’s voice pipeline or an online AI service, with the ESP32 streaming audio to it.

How do I make a Home Assistant voice satellite with an ESP32?

Flash the board with ESPHome using a voice assistant configuration (ready-made ones exist for M5Stack Atom devices and many others), add it to Home Assistant, and set up an Assist pipeline with speech-to-text and text-to-speech. An ESP32-S3 with PSRAM can run the wake word on the device itself.

How do I wire an INMP441 and MAX98357A to an ESP32?

Each needs a bit clock (SCK/BCLK), a word select (WS/LRC) and a data line (SD for the mic, DIN for the amp). They can share the clock and word-select pins on one I2S port or use two ports. Power the INMP441 from 3.3 V and the MAX98357A from 5 V, and connect all grounds together.

Why does my voice assistant not hear me while music plays?

Without acoustic echo cancellation, the microphone hears the speaker louder than your voice. Boards with a voice processor such as the XMOS XVF3800 remove the speaker’s output from the mic signal, so you can interrupt playback. Lowering volume or moving the mic away from the speaker helps on simpler boards.

Can the original ESP32 play Bluetooth audio?

Yes. The original ESP32 supports Bluetooth Classic with the A2DP profile, so it can act as a Bluetooth speaker or audio source. The ESP32-S3, C3, C5 and C6 only support Bluetooth Low Energy, so they can’t receive standard Bluetooth music streams. Use an original-ESP32 board such as the ESP32-Audio-Kit for Bluetooth speakers.