aria taste-steering directions

Activation-steering directions for Stable Audio 3, from the paper A Quantized Native Runtime for On-Device Semantic Audio Generation (IEEE IS2 2026). They bias generated music toward a basic taste (sweet, sour, bitter, salty, spicy), the crossmodal effect known as sonic seasoning.

They are meant for aria, the dependency-free C/CUDA runtime for Stable Audio 3, which applies them inside the generation graph with no measurable overhead. Demo page with audio examples: https://matteospanio.github.io/aria/.

No model weights are included. You need the Stable Audio 3 checkpoints, which are gated and distributed by Stability AI under the Stability AI Community License.

Contents

folder	model	site	vectors
small/	small-music (1024-D)	DiT residual stream	sweet, sour, bitter, salty, spicy
medium/	medium (1536-D)	DiT residual stream	sweet, sour, bitter, salty, spicy
sites/	small-music	audio latent (256-D), text conditioning (768-D)	sweet, sour

Each .atns file is one float32 vector in aria's tensor format. dirs_meta.json lists, per taste, the DiT block the vector belongs to (layer), the mean residual norm it was calibrated with (norm) and its dimension.

Each residual vector is a difference-in-means of DiT block outputs between the top- and bottom-rated clips of the norm-sonic-seasoning dataset (377 clips with human taste ratings), normalized to unit length and multiplied by the block's mean residual norm. The scale you pass to aria is therefore the paper's α, in units of the typical residual magnitude. The sites/ vectors are unit vectors; their scale was swept empirically.

Usage

hf download matteospanio/aria --local-dir directions

./aria -m models/small-music -p "lo-fi hip hop beat" --seed 22 -d 10 -s 8 \
       --steer residual:19:directions/small/sour.atns:0.1:0-7 -o sour.wav

The spec is site:layer:file.atns:scale:first-last step[:add|project]. The site is residual, latent or cond (text conditioning). The layer is ignored for latent and cond.

Operating points from the paper

The usable range is narrow. Past it, the target taste score keeps rising while independent checks (CLAP similarity, FAD) show the audio degrading. Every point below was selected with those checks, never with the taste score alone.

model	taste	spec	note
small-music	sweet	residual:16:small/sweet.atns:0.3:0-7	genuine, widest clean window
small-music	sour	residual:19:small/sour.atns:0.1:0-7	genuine
small-music	sour	residual:19:small/sour.atns:1:0-7:project	same shift, lower FAD
small-music	sour	residual:19:small/sour.atns:0.2:4-7	late steps only: about half the damage
small-music	bitter	residual:19:small/bitter.atns:0.1:0-7	genuine
small-music	salty, spicy	residual:19:…:0.1:0-7	weak even when clean
small-music	sour	latent:0:sites/sour_latent.atns:1:0-7	genuine at the latent site
small-music	sweet	latent:0:sites/sweet_latent.atns:0.5:0-7	weak at 0.5; inverts at larger scales
small-music	sweet, sour	cond:0:sites/*_text.atns	fails at every scale tested (0.5–32)
medium	sweet	residual:21:medium/sweet.atns:0.15:0-7	genuine
medium	sour, bitter	residual:23:medium/<taste>.atns:0.1:0-7	genuine
medium	salty	residual:5:medium/salty.atns:0.1:0-7	weak
medium	spicy	residual:23:medium/spicy.atns:0.1:0-7	weak

Taste control is genuine only for sweet, sour and bitter, and only inside these windows. Treat it as a bounded proof of concept, not general taste control.

Citation

@misc{spanio2026quantizednativeruntimeondevice,
  title         = {A Quantized Native Runtime for On-Device Semantic Audio Generation},
  author        = {Matteo Spanio and Antonio Rodà},
  year          = {2026},
  eprint        = {2607.08526},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2607.08526},
}