AWS DeepComposer concepts and terminology - AWS DeepComposer

AWS DeepComposer concepts and terminology

AWS DeepComposer builds on the following concepts and uses the following terminology.

AutoregressiveCNN Bach

A pretrained autoregressive convolutional neural network (AR-CNN) model that is available in the AWS DeepComposer music studio. This model has been trained on a dataset containing only Bach chorales. It uses the U-net architecture.

autoregressive convolutional neural network (AR-CNN)

A generative AI technique that edits your input melody. This technique uses a U-Net architecture and was trained on a collection of chorales by Johann Sebastian Bach. The AR-CNN technique works by detecting notes that sound missing or out place based on the training dataset. The identified notes are then replaced with notes that the model thinks would fit the distribution of notes learned during training.

autoregressive model (AR model)

A linear model often used with time-series data, where previous values are used to predict a current or future value.

AWS DeepComposer keyboard

Also known as the AWS DeepComposer physical keyboard or hardware keyboard. You can connect, or link, the keyboard to a computer that has access to the AWS DeepComposer console. You can use the linked keyboard to play and record short melodies that are fewer than eight bars. Then use a recorded melody with a supported AWS DeepComposer generative AI technique.


Sequences of notes, melodies, harmonies, and rhythms that make up a musical work.

In AWS DeepComposer, a composition also represents a saved piece of music. Each time you use a generative technique to perform inference, your output is automatically saved as a composition.

computer vision (CV)

A field of machine learning that is focused on creating models that understand the visual world by interpreting images and videos. Computer vision can be used for tasks like objection identification and video tracking.

convolutional neural network (CNN)

A type of neural network commonly used for image-recognition and video-recognition tasks. A CNN filters and summarizes the data during training to learn patterns. This mathematical process is called convolution.


A process that takes an input vector, commonly a feature matrix, and transforms the vector into an output. A decoder can be used with an encoder. AWS DeepComposer uses an encoder-decoder network for some generative AI tasks.


A term used to describe a structured collection of data.

The data can be thought of as simply an example or observation.

Before model training, datasets are often split into subsets of data to help ensure accurate results. For more information about two common subsets see, testing dataset and training dataset.


A classifier model, one part of a generative adversarial network (GAN). The discriminator classifies data as real or as generated. In AWS DeepComposer, this model tries to determine if a generated piano roll looks like a real image or an artificially created image. The discriminator is trained on real data.

drum in pattern

A ratio of the total number of notes in a drum track to a predetermined beat popular in 4/4 time.

empty bar rate

The ratio of empty bars to the total number of bars.


A process that translates data from one format to another. AWS DeepComposer uses multiple encoders. For example, it uses an encoder to translate an input MIDI file into a piano roll image. Often, an encoder is used with a decoder.


One complete pass through the training dataset by the neural network. For example, if you have 10,000 music tracks in the training dataset, one epoch represents one pass through all 10,000 tracks. The number of epochs required for a model to converge varies based on the training data. An iterative algorithm converges when the loss function stabilizes.

generative adversarial network (GAN)

Two neural networks that consist of a generator and a discriminator. In AWS DeepComposer, the generator learns to compose music that sounds as realistic as possible with feedback from the discriminator. The discriminator treats the generator's output as sounding as unrealistic as possible while holding the input training sample as the ground truth. Training proceeds, with the generator searching for its network weights by minimizing the chances that its generations differ from the training samples. The discriminator searches for the weights of both networks by minimizing the chances it misjudges the training samples as realistic compositions and the generator's outputs as unrealistic compositions. The efforts made by both the generator and the discriminator push the generator's network weights in opposite directions. This is the essence of the adversarial role that the discriminator takes against the generator. Learning proceeds until equilibrium is reached, where the generator improves its output to such a degree that the discriminator can no longer distinguish between the generated composition and training samples.


Algorithm-dependent variables that you can control. You can tune hyperparameters to find the best fit for a specific problem that you are trying to model.


Predictions generated by a trained model.

in scale ratio

The ratio of the average number of notes in a bar of music that are in the key of C, to the total number of notes in a bar of music.

learning rate

A hyperparameter used for training neural networks. The learning rate controls how much the weights and biases are updated during training.

loss function

Evaluates how effective an algorithm is at modeling the data. For example, if a model consistently predicts values that are very different from the actual values, it returns a large loss. Depending on the training algorithm, more than one loss function might be used.


The final output created while training with a machine learning algorithm. The algorithm used to train the model finds patterns in the training data uses those patterns to map the input data attributes to the target (the answer that you want to predict). The algorithm outputs a machine learning model that captures these patterns.


A generative adversarial network (GAN) architecture built specifically to generate music. Like other GANs, MuseGAN is made of both a discriminator and a generator that use a CNN. The MuseGAN architecture is available in AWS DeepComposer. To learn more about the MuseGAN architecture, see Creating the MuseGAN architecture.

Music studio

A component of the AWS DeepComposer console that provides access to a console-based keyboard for playing, recording, and composing music. With Music studio, you can try out AWS DeepComposer before purchasing an AWS DeepComposer physical keyboard. You don't need to train a model to use Music studio.

neural network

Also known as an artificial neural network. A collection of connected units or nodes that are used to build an information model based on biological systems. Each node is called an artificial neuron. An artificial neuron mimics a biological neuron in that it receives an input (stimulus), becomes activated if the input signal is strong enough (activation), and produces an output based on both. Neural networks are widely used in machine learning because an artificial neural network can serve as a general-purpose approximation to any function.

piano roll

A type of music roll that encodes a musical score in a manner that is machine readable. In AWS DeepComposer, they are used to train both the GAN models and the AR-CNN model. In each image, the horizontal axis represents time and the vertical axis represents pitch.


The frequency of a sound, often judged to be high or low. In the AWS DeepComposer Music studio, you can adjust the pitch of a sample, imported, or custom recorded input melody.

pitches used

A metric that captures the average number of notes in each bar.

polyphonic rate

The ratio of the number of time steps where the number of pitches being played is greater than the threshold number of allowable time steps.

Rhythm assist

An AWS DeepComposer music editing tool that automatically corrects the timing of musical notes in your input. Use Rhythm assist when your melody contains the right notes but isn't consistently in time with the beat. Rhythm assist is available in classic music studio studio experience.


How fast music is played. Music typically follows a certain beat or meter, which drives the rhythm of the notes played. The speed of this beat is measured in beats per minute. A higher number of beats per minute corresponds to a faster tempo, or playback speed.

testing dataset or test set

Part of a dataset used evaluate the accuracy of the model trained using a training dataset.


A generative AI technique that uses attention to understand the relationship between notes in a musical composition. To train the model, the musical data are converted into tokens. These tokens are used to represent a musical event in a given piece of music. During inference, the Transformers technique will extend your input track by up 30 seconds.


A model based on the Transformer architecture. Compared with the traditional Transformer architecture, this model can better understand long-term dependencies and has decreased latency times.

training dataset

Part of a dataset used while training a machine learning model.


A generative adversarial network (GAN) architecture built originally for image recognition tasks. The U-Net is a CNN. It's named for its U-like shape, in which the layers on the left side can pass information to the layers on the right side without passing through the entire neural network.

update ratio

A hyperparameter that controls the number of model weight updates to the discriminator per update to the generator. A lower update ratio makes a stronger discriminator that can provide more accurate and useful information to the generator. The lower update ratio, however, increases training time.

virtual keyboard

Also known as the AWS DeepComposer console-based keyboard. The virtual keyboard is available in the AWS DeepComposer Music studio. The virtual keyboard runs on any computer or mobile device that is connected to the AWS Cloud. You can use the virtual keyboard to play and record a short melody. You can then use the recorded melody with a supported AWS DeepComposer generative AI technique to create new musical compositions.