xaitalk/audionet-audiomnist

AudioNet — AudioMNIST 10-class spoken-digit classifier

A small 1D-CNN trained from scratch on AudioMNIST (10-class spoken digit recognition, 8000-sample mono audio at 8 kHz). Released as a real-pretrained-weights demo for xaitalk's cross-framework XAI on 1D-conv audio architectures.

This is not a state-of-the-art audio model. It exists so XAI demos on speech / audio work on a realistic-but-small architecture (no HuggingFace 100M-param dependency, no datasets-streaming workflow).

Files

File	Format	Size
audionet_audiomnist_full.keras	full Keras model (architecture + weights)	~10 MB
audionet_audiomnist.weights.h5	Keras weights only	~10 MB

Architecture

Property	Value
Input	1D waveform, length 8000 (= 1 second at 8 kHz)
Conv channels	64 → 128 → 256 → 512
Dense	1×256
Output	10-class logits (digits 0-9)
Total params	~3 M

Training: 30 epochs on the AudioMNIST training split (Becker et al. 2018). Final test accuracy: 48.1% (training script is in the xaitalk repo at examples/training/train_audionet_audiomnist.py).

The accuracy is intentionally moderate — the model is small enough to train quickly and large enough that the per-layer attribution maps are visually meaningful, which is what the XAI demo needs.

Cross-framework verification

These weights are validated by the xaitalk method × architecture matrix on the audio benchmark (Conv1D, 20 methods):

Methods	Passing at r ≥ 0.95	Min(min_r)	Verified
20	20/20	0.9973	2026-05-09

Includes gradient family (9), LRP variants (8), DeepLIFT, smoothgrad family (smoothgrad, smoothgrad_sq, vargrad). Full results: xaitalk method matrix.

Usage

from xaitalk.hub import ensure_model

# Full Keras model (architecture + weights):
keras_path = ensure_model('audionet/keras')
# Weights-only (you supply the architecture):
h5_path    = ensure_model('audionet/h5')

# Load full model
import tensorflow as tf
model = tf.keras.models.load_model(keras_path)

# Run XAI
import xaitalk
import numpy as np
x = np.random.randn(1, 8000, 1).astype(np.float32)
expl = xaitalk.explain(model, x, method='lrp_epsilon', target_class=5)

Training data

AudioMNIST — a spoken-digit recognition dataset by Sören Becker et al. (Free Spoken Digit Dataset). 30000 recordings of digits 0-9 spoken by 60 speakers.

License

Apache 2.0. AudioMNIST is released under MIT (see upstream).

Citation

Original AudioMNIST dataset:

@misc{becker2018audiomnist,
  author = {Becker, Sören and Ackermann, Marcel and Lapuschkin, Sebastian
            and Müller, Klaus-Robert and Samek, Wojciech},
  title  = {Interpreting and Explaining Deep Neural Networks for
            Classification of Audio Signals},
  year   = {2018},
  eprint = {1807.03418}
}

xaitalk infrastructure:

@software{paul2026xaitalk,
  author = {Paul, Alexander},
  title  = {xaitalk: Cross-Framework Explainable AI Library},
  year   = {2026},
  url    = {https://xaitalk.com}
}

Links

• xaitalk website: https://xaitalk.com
• Framework GitHub: https://github.com/alexanderfpaul/xaitalk-framework
• Training script: examples/training/train_audionet_audiomnist.py
• AudioMNIST: https://github.com/soerenab/AudioMNIST