Music generation
Notes are tokens too: predict the next note like the next word.
Music has both pitch and duration, so the Transformer gets two input/output streams. Polyphonic music needs a tokenisation choice: grid (fixed time slots) or event-based (a sequence of instructions). MuseGAN instead uses convolutions to generate multi-track scores like images.
The example uses the simplest autoregressive model — a bigram — over note:duration tokens and composes three short melodies.
| Approach | Idea |
|---|---|
| Transformer (note + duration) | Two embeddings combined; predicts next note and duration |
| Grid tokenization | Equal time steps, all voices read as one stream |
| Event tokenization | Instructions: note on, wait, note off |
| MuseGAN | Treats the score as an image (piano roll) and uses a GAN |
Key takeaways
- Music models live or die by their tokenisation
- Models learn keys and scales on their own just by predicting the next note
import numpy as np
rng = np.random.default_rng(1)
# A tiny corpus of melodies as (note, duration) tokens
songs = [
"C4:1 E4:1 G4:1 E4:1 C4:2", "C4:1 D4:1 E4:1 G4:1 E4:2",
"E4:1 G4:1 A4:1 G4:1 E4:2", "G4:1 E4:1 D4:1 C4:1 C4:2",
]
tokens = [s.split() + ["END"] for s in songs]
vocab = sorted({t for song in tokens for t in song} | {"START"})
ix = {t: i for i, t in enumerate(vocab)}
# Bigram model p(next | current): the simplest autoregressive music model
counts = np.zeros((len(vocab), len(vocab)))
for song in tokens:
for a, b in zip(["START"] + song, song):
counts[ix[a], ix[b]] += 1
probs = counts / np.maximum(counts.sum(1, keepdims=True), 1)
def compose(max_len=10):
cur, out = "START", []
for _ in range(max_len):
cur = vocab[rng.choice(len(vocab), p=probs[ix[cur]])]
if cur == "END": break
out.append(cur)
return " ".join(out)
for i in range(3):
print(f"melody {i + 1}: {compose()}")Output
melody 1: E4:1 G4:1 A4:1 G4:1 E4:1 G4:1 E4:2 melody 2: E4:1 C4:2 melody 3: E4:1 D4:1 E4:1 D4:1 C4:1 C4:2
- #GenerativeAI
- #DeepLearning
- #PythonສຳລັບຄົນLaos
- #AI
- #ບອກຕໍ່ຄວາມຮູ້