Component guide · Intermediate · 3 min read
Voice Recognition and Recording Modules (ISD1820, LD3320)
By Geekobot Learning Team · Updated 24 Oct 2026
On this page
Voice modules come in two kinds. Record and playback modules such as the ISD1820 store a short message, about 10 to 20 seconds, and play it at the press of a button or a pin signal. Voice recognition modules such as the LD3320 or Elechouse V3 recognise a small set of trained or programmed commands offline and send a result over UART. They suit simple voice control, but cannot understand free speech.
Key specs at a glance
- ISD1820
- 8 to 20 s recording, onboard mic, REC, PLAYE and PLAYL inputs
- Voice Recognition V3
- Up to about 80 trained commands, 7 active at once, UART
- LD3320
- Keyword list set in code (pinyin-based on many boards), SPI or parallel
- Supply
- 3.3 V or 5 V depending on module
- Language
- Trained modules work with any language; keyword chips depend on the model
- Accuracy
- Best in quiet rooms, close to the mic, with the trained speaker
Record and playback (ISD1820)
Power it from 3 to 5 V and connect a small 8 Ω speaker. Hold REC to record, press PLAYE to play the whole message or hold PLAYL to play while held. Drive PLAYE from a GPIO pin (pulse it high) to play a message when a sensor triggers, such as a PIR greeting or a parking warning. For multiple messages and better quality, use an MP3 module instead.
Voice recognition modules
Trained modules (such as the Elechouse Voice Recognition V3) record a few samples of each command you speak, then compare new speech against them. They work in any language, but best for the person who trained them.
Keyword modules (such as the LD3320) recognise words from a list you program, without training, but supported languages depend on the chip.
Both send a command number over UART or SPI. Your code then switches a relay, turns a servo or plays a sound. Use a short wake word followed by the command to reduce false triggers.
When to use something more powerful
For more commands, better accuracy or natural speech, use an ESP32-S3 with Espressif's offline speech models, or a Raspberry Pi with an open-source speech engine. Cloud assistants understand the most but need internet and send audio off the device. For an I2S microphone input, see that guide.
Common mistakes to avoid
- Training commands in a quiet room and using them in a noisy one.
- Choosing commands that sound alike.
- Expecting a keyword module to understand full sentences.
- Powering the speaker from a weak supply and getting resets.
- Using recognition for safety-critical control.
Frequently asked questions
Can a voice module understand Hindi?
Trained modules like the V3 work with any language you train. Keyword chips depend on their language model.
How long can the ISD1820 record?
About 10 seconds by default; changing the sample-rate resistor gives up to about 20 seconds at lower quality.
Do voice modules need internet?
No. These modules recognise commands offline.
How many commands can I use?
Usually a handful active at once; the V3 stores about 80 but listens for 7 at a time.
Why does it trigger by mistake?
Background noise or similar-sounding words. Use distinct commands and a wake word.
Parts for this guide
Terms in this guide
Projects that use it
Put this into practice: 2 step-by-step builds in our project library use parts from this guide.
Keep learning
Sources
Written by the Geekobot team from manufacturer datasheets. Always check the datasheet for the exact part and variant you buy.


