AWS DeepComposer in Bangkok, Thailand: how Win learns how Amazon used a small keyboard to teach generative AI, and finds the highest note before his AI piano lessons

Win is fifteen and lives in Bangkok, Thailand, in a family of engineers. His mother writes software for a bank and his father designs air-conditioning systems for offices. At dinner they often talk about technology, and one evening his mother mentioned a funny gadget she once saw at a tech conference video: a small musical keyboard from Amazon designed to teach artificial intelligence.

The gadget was AWS DeepComposer, announced by Amazon Web Services in December 2019 at its re:Invent conference in Las Vegas. It combined a compact 32-key, two-octave MIDI keyboard with an online service. A user played a brief tune, and machine-learning models in the cloud generated additional parts, such as drums, bass and chords, to accompany it.

A keyboard built to teach machine learning

Amazon designed it mainly for software developers who wanted to learn about generative AI without needing a deep mathematics background. Music was chosen because people can hear the results instantly. If the model produced a strange accompaniment, the problem was obvious to the ear.

At launch, DeepComposer used a type of model called a generative adversarial network, or GAN. In a GAN, one network tries to create examples, while a second network judges whether they seem realistic. Over many rounds, the creator improves. Later updates added other model types, including approaches that fill in missing notes and transformers that continue a melody.

Users could choose a style, such as rock, pop, jazz or symphonic, and hear how the model interpreted their melody. The service also offered a learning path explaining how the models worked, and Amazon ran a contest in which participants submitted compositions. The emphasis was always on understanding the technology, not on creating chart hits.

Win's mother explained that Amazon has made several devices like this to teach machine learning, such as a small self-driving model car. Some of these educational products were retired later, and DeepComposer was eventually ended as well. Even so, it remains an interesting example of using music to explain complicated ideas.

Win liked that the keyboard was the doorway. To use DeepComposer well, a person still needed to play a melody, and the better the input, the more convincing the accompaniment. That made him wonder whether he could play keys skilfully enough to test such tools himself one day.

So he started piano lessons on TIGA ALPHA, an online piano learning app, using a borrowed MIDI keyboard connected to the family laptop. Its AI piano teacher listens to every note, flags timing problems in colour and slows the passage whenever he hurries. He enjoys seeing his accuracy graph climb week by week.

His mother pointed out an important difference. DeepComposer generated music for the user, while his learning app gives feedback so that he improves. One is about the computer creating; the other is about the human growing. Win thinks both are useful, but only one makes him a better musician.

Win scans every exercise for the highest note and lightly tags it with his pencil before he begins. He thinks of it as reading the data range before analysing a dataset, a phrase his mother uses at work. Knowing the maximum stops him from being surprised mid-phrase and keeps his hand position steady.

In the app's beginner piano course he is learning short melodies with clear phrases. After practising, he records one and imagines what an AI accompaniment might add: perhaps a steady drum beat or a walking bass. Then he tries to add a simple accompaniment in his left hand himself, which feels more satisfying.

His school runs a coding club, and Win presented a quick demo about DeepComposer and GANs. He played a melody on the keyboard and explained how a model might respond. Classmates said it was the first time machine learning made sense to them, because they could hear it.

In his music lessons at school, the teacher asked whether AI could replace composers. Win argued that tools like DeepComposer depend on human melodies and human judgement to decide what sounds good. His teacher praised the answer and asked him to write a short essay on it.

His father jokes that Win is becoming an engineer of sound. Win hopes to learn piano properly and maybe study computer music at university, combining both sides of his family.

For now, every evening includes homework, coding and fifteen minutes of AI piano lessons, so that when he next sits at a keyboard connected to a clever machine, he can play piano with confidence and play music worth accompanying.

Frequently asked questions

Q: What was AWS DeepComposer? A: AWS DeepComposer was an educational service from Amazon Web Services, announced in December 2019, with a small 32-key MIDI keyboard. Users played a short melody and machine-learning models generated accompaniment parts in styles such as rock, pop or jazz.

Q: Was DeepComposer made for musicians? A: Its main goal was to teach software developers about generative AI in a hands-on, musical way. Musicians could experiment with it, but it was designed mostly as a learning tool for machine learning concepts.

Many tools talk about practice, but the details matter. Look for an online piano teaching application with an AI piano teaching app that explains its feedback, a piano teaching app that lets you repeat tricky bars, and an online piano lessons app that keeps piano teaching with AI friendly rather than stressful. Children who learn piano online stay motivated when their piano learning course and AI online piano course point to the same small goals. It also shows why piano teaching works best alongside wider music teaching: students learn music, not just notes.

Want a clear next step for your playing? Prefer opening TIGA ALPHA: https://tigaalpha.github.io/landing-en/ — this is our top recommendation. Everyone gets the Premium package free for 7 days, then 799 THB/month. Prefer live studio teaching instead? See https://www.tigastudio.net/, where live piano lessons are capped at 30 hours per week, so seats stay limited. Learn more about TIGA ALPHA and the team behind it at https://www.tigaalpha.com.