Listening demos

TASTE: taxonomy-aware symbolic piano generation from text

Piano generators trained on MIDI tokens with a three-tier corpus (large public → curated → high-quality) and conditioned on taxonomy attributes, key, tempo, section plan and free-text imagery. Each AR page shows four hand-written imagery prompts with three random seeds each; the DiT page keeps its five earlier prompts. Each piece is rendered to audio with a synced piano-roll.

TASTE (AR)

TASTE (AR) — high-quality stage

final checkpoint

0.89B decoder pretrained on piano MIDI, then conditional stages on the 371K large-public tier, the 51K curated tier and the 10K high-quality tier.

  • Resolution 25 ms, 4,096-token context
  • Imagery encoder Qwen3-Embedding-4B (2560-d)
  • Sampling T 0.98 · top-p 0.98
Open demo →

TASTE (AR) — best-of-8 by CLaMP 3 selection

final checkpoint

0.89B generator, high-quality-stage checkpoint; four prompts, eight samples each, top three shown ranked by the CLaMP 3 selection score (text cosine + attribute-prototype cosine) minus a key / tempo read-back penalty.

  • Resolution 25 ms, 4,096-token context
  • Imagery encoder Qwen3-Embedding-4B (2560-d)
  • Sampling T 0.98 · top-p 0.98, best-of-8
Open demo →

Earlier stages

TASTE (AR) — curated stage

earlier stage

0.89B decoder, pretrained on piano MIDI, large-public-stage training on 371K pieces, then curated-stage training on 51K annotated pieces.

  • Resolution 25 ms, 4,096-token context
  • Imagery encoder Qwen3-Embedding-4B (2560-d)
  • Sampling T 0.98 · top-p 0.98
Open demo →

TASTE (AR) — large-public stage

earlier stage

0.89B decoder, pretrained on piano MIDI, then large-public-stage training on 371K pieces with weakly labelled attributes.

  • Resolution 25 ms, 4,096-token context
  • Imagery encoder Qwen3-Embedding-4B (2560-d)
  • Sampling T 0.98 · top-p 0.98
Open demo →

Ablation

AR — skipping the large-public stage

ablation

Same architecture and recipe, but the conditional training goes pretrain → curated → high-quality without the large-public stage. Same four prompts and seeds as the high-quality page for a card-by-card comparison.

  • Resolution 25 ms, 4,096-token context
  • Imagery encoder Qwen3-Embedding-4B (2560-d)
  • Sampling T 0.98 · top-p 0.98
Open demo →

TASTE (DiT)

TASTE (DiT) — high-quality stage, cfg 3.0 vs no guidance

latent DiT

397M DiT over the window-VAE latent space (two 128-d latents per second, 192 s canvas, rectified flow), trained through the same three tiers; samples from the final high-quality stage. The DiT takes the section plan and piece length as time-aligned inputs, so each prompt borrows them from a high-quality validation piece with the same emotion and genre. Each card carries six tabs: three seeds at classifier-free guidance 3.0 and the same seeds with guidance off.

  • Latent 2 per second, 128-d · 192 s canvas
  • Imagery encoder Qwen3-Embedding-4B (2560-d), cross-attention
  • Sampling EMA weights · CFG 3.0 and 1.0 · 50 Euler steps, seeds 101–103
Open demo →

Codec

Window VAE reconstruction — 12 pieces, 4 per tier

codec check

The MuseTok-style window VAE (one 128-d latent per second, 12-layer encoder/decoder) that TASTE (DiT) generates into. Each card plays a piece's original token stream next to its reconstruction from the latents, decoded free-running per 20 s chunk with a token-grammar mask; note F1 0.94 against the original at ±50 ms, velocity 99.6 % and duration 99.3 % agreement.

  • Latent 1 per second, 128-d
  • Pieces 4 large-public · 4 curated · 4 high-quality at note-density quantiles
  • Decoding constrained greedy, chunked 20 s
Open demo →