---
license: mit
tags:
  - multimodal
  - audio
  - image
  - embeddings
---

# Cross-modal associative embeddings (MNIST + AudioMNIST)

Frozen encoder embeddings for a cross-modal (image <-> audio) associative
memory experiment. Generated by the `associative-learning` project; see
`MANIFEST.json` for exact provenance.

## Contents

| File | Rows | Dim | Description |
|---|---|---|---|
| `image_train.npz` | 60,000 | 768 | ViT pooled embeddings of MNIST train images |
| `image_test.npz` | 10,000 | 768 | ViT pooled embeddings of MNIST test images |
| `audio_train.npz` | 24,000 | 768 | AST pooled embeddings of AudioMNIST train clips |
| `audio_test.npz` | 6,000 | 768 | AST pooled embeddings of AudioMNIST test clips |

Each `.npz` holds `embeddings` (N, 768) plus metadata:

- `label` — digit class (0-9)
- `key` — stable source identifier
- audio files additionally carry `speaker_id` and `gender` (0=male, 1=female)

## Encoders (frozen, never fine-tuned)

- **Images:** `google/vit-base-patch16-224` (ViT-base)
- **Audio:** `MIT/ast-finetuned-audioset-10-10-0.4593` (AST)

Audio is resampled to 16 kHz; the AST feature extractor computes the log-Mel.

## Split

The upstream splits leak speakers, so the split here is rebuilt:
**speaker-disjoint and gender-balanced**. AudioMNIST's shipped test set is only
1/12 female (8%), which cannot support a gender-generalization claim.

## Pairing caveat

MNIST and AudioMNIST share a *label*, never an instance — no image of a "3" was
spoken as that same "3". Association here is **label-mediated**, not
instance-level. Pairing is a Cartesian product, so every image of a digit may
be paired with every clip of that digit, in both directions.