AI glossary · Basic · Language models · Speech & audio
RNN
An RNN (recurrent neural network) is an artificial neural network that reads ordered data such as sentences or speech one piece at a time, carrying what it has seen so far forward as 'memory' for the next decision.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Models & generation
- Fields
- Language models, Speech & audio
- Added
- 2026-10-08
How it works
An RNN receives input one item at a time (for example, one word at a time). At each step, it combines the current input with the memory so far to create a new memory, called the hidden state. Because this memory keeps being passed on to the next step, it is called 'recurrent'. Based on the memory at the final step, it produces a result, such as predicting the next word or classifying the sentiment of a sentence.
Why it matters
Ordinary neural networks have a fixed input length and do not specifically account for order. RNNs can process sentences or sounds of varying lengths in order, so they were once the basic tool for translation, speech recognition and text generation. Improved variants such as LSTM and GRU were developed to reduce the problem of memory fading quickly, and became widely used. As the starting point for handling sequential data with neural networks, RNNs are still taught as a foundational concept.
Things to know
RNNs have the weakness that information from the beginning of a sentence is poorly carried forward as the sentence gets longer. They also have to process words one at a time in order, which makes parallel computation on GPUs difficult and training slow. Because of these limits, most of today's LLMs use transformers. However, architectures that efficiently revive the idea of 'processing in order while carrying memory forward', such as recent state space models, are drawing attention again.
Example
RNN-family models (especially LSTM) were widely used for next-word suggestions in early smartphone keyboards, speech recognition in voice assistants, online translators and more. Many have since been replaced by transformers, but they are still used for lightweight analysis of time-ordered data such as sensor data or stock prices.