How it works

The original decoder took the single most likely next note, every time. That sounds reasonable and it is quietly fatal: a deterministic decoder can only ever walk one path out of a given passage. The moment the model saw a context it had seen before, it played the same continuation, forever. What sounded like a model with nothing to say was really a model never being asked.

The fix is to sample from the distribution rather than collapse it. Temperature controls how much: low values sharpen toward the model's favourite note, high values flatten toward surprise.

Calibrating it was the interesting part, because this model is unusually sure of itself. Its next-note distribution has a mean top probability of 0.95 and an entropy of 0.21 nats — far sharper than a language model. The likeliest note already holds more probability than any sensible nucleus threshold, so the textbook defaults borrowed from text generation truncate to a single token and silently reproduce greedy decoding. The dial does nothing, and you would never know.

Below about 1.3 the decoder is greedy. Above about 2.1 it is drawing from a flat tail of two thousand notes, which is noise. The usable band is 1.4 to 2.0, and that is the range the creativity slider spans — not an arbitrary 0 to 1.

What that buys, averaged over four seeds at 160 notes each. Repetition is the share of four-note windows already played; corpus KL measures distance from the training set's pitch distribution. Lower is better in both.

The second row is the point. Those are the sampling defaults you would copy from any text-generation tutorial, and they are indistinguishable from greedy decoding — the difference sits well inside the variation between seeds. The dial would have been connected to nothing.

One honest caveat. Sampling diverges from the greedy path slowly: at 120 notes it changes only a fiftieth of the model's decisions, but by 300 notes it has changed most of them. Sampling matters more the longer the piece runs.

The network takes a window of past notes and nothing else. It has no input for key, no input for mood, and training a model that did would need labelled data nobody has. So conditioning happens at the last possible moment instead: every note in the vocabulary gets a score, and that score is added to the model's own opinion just before a note is picked.

Ask for C major and notes outside the scale get pushed down — in proportion to how far outside they are, so a chord with one foreign note is discouraged less than one with three. The push is deliberately gentle, a few nats. A heavier hand would win every argument with the model and flatten the music into scales.

Mood presets bundle a temperature, a scale, a register and a note density. They exist because “melancholic” is easier to ask for than temperature 1.45, natural minor, centred on G3, favour long notes.

The tightened spread matters as much as the raised average. Left alone, which key you get is luck of the seed. Asked, you get the key you asked for.

Comparing architectures is only meaningful when nothing else moves. All three share a dataset, a tokenizer and a decoder, so a difference in output is a difference in architecture rather than in plumbing.

That middle row is worth staring at. The original network ends by flattening a 100×512 sequence into a single dense layer, which costs 175 million parameters on the output layer alone. The transformer reaches comparable capacity with 46× fewer, because it never builds that matrix. Whether it also sounds better is not a question a parameter count can answer — which is what the arena is for.

Validation loss tells you how well a model predicts the corpus. It does not tell you whether a person would want to listen to it. Cross-entropy punishes a beautiful passing tone for the crime of not appearing in Chopin.

So the arena asks a person. Two architectures compose the same brief — same length, key, temperature and random seed, so the architecture is the only thing that differs — and you hear them without being told which is which. Elo ratings update from your vote, the same system chess uses, because votes arrive as pairwise comparisons and a new model can join without re-running every earlier matchup.

Blinding is enforced on the server rather than in the interface: the model name is stripped from the response, kept out of the progress log, and the metadata file that would otherwise sit next to each clip is not written. It would have been reachable by guessing a URL.

Two cheap statistics run alongside the human vote. Repetition is the share of four-note windows the decoder has already played. Corpus KLcompares the piece's pitch-class distribution against the training set. Neither replaces listening. Both are reproducible, which listening is not.

MAiSTRO composes notes, and a piano renders them. That is exactly right for classical piano and exactly wrong for music that lives in timbre. There is no arrangement of piano notes that means “dusty Rhodes through a tape delay.”

MusicGen predicts compressed audio instead of notes, so it can render any genre it was trained on — at the cost of never producing a score you can edit. It runs on the Python backend and is an optional install. Magenta's MelodyRNNdoes the same job as MAiSTRO's LSTM, continuing a melody one note at a time, but downloads to your browser and runs on your own GPU. Together they bracket the question of where a model should live.

This site runs on a free tier. Vercel's Python runtime is 3.12 with a 500MB ceiling on a deployed function — and TensorFlow is 877MB unpacked, and needs Python 3.10 or older. The training backend cannot go there at any size.

The way through is a distinction worth internalising: TensorFlow is only needed to train. Running a trained model forward is a dozen matrix multiplications over a few million numbers. Nothing about that requires a deep-learning framework.

So the deployed API reimplements the transformer's forward pass in plain NumPy and reads its weights from a 7MB file. Writing MIDI moves from music21 (111MB) to mido (under 1MB). What ships is fastapi, numpy and mido.

A rewrite is only worth anything if it computes the same function. The NumPy path is checked against Keras on identical weights: the largest disagreement anywhere in the output distribution is 4.2 × 10⁻⁷, the two always pick the same next note, and a 300-note piece takes 9.9 seconds against a 300-second limit. The training notebook re-runs that comparison and refuses to export weights that fail it.

Which model gets deployed follows from the arithmetic. The original LSTM keeps 97% of its 178.8M parameters in one output layer — 357MB even at half precision. The transformer's 3.9M parameters (256 dimensions, 4 heads, 4 layers) fit in 7MB. It is trained on a free Colab T4, where an epoch takes about two minutes instead of the seventeen it takes on a laptop.

A serverless function is stateless with a read-only disk, so the deployment keeps what fits and is honest about the rest. Generation answers in one request, returning the MIDI inline rather than writing a file. Magenta was already running in your browser. The library, the arena, training and MusicGen need a filesystem, two trained models, hours, and 511MB of torch respectively — they stay local, and the interface says so rather than failing at you.

Every dependency here earns its place by removing a specific problem.

The transformer ships untrained. Its architecture, training path and arena integration are tested end to end, but nobody has run the epochs, so the comparison above is a parameter count and not a quality claim.

Conditioning is steering, not learned control. A strong bias argues with the model rather than collaborating with it, which is why the penalties are small.

The arena's ratings mean little until a few dozen votes are in. One listener cannot invert a leaderboard, but they can certainly wobble it.