Spoken-digit recognition

A convolutional neural network for spoken-digit recognition, developed for the Kaggle iML challenge at ULiège. The model combines 2D and 1D convolutions to classify digits 0–9 and other/non-digit from 13 MFCC coefficients over 32 time steps.

Install

python -m pip install -r requirements.txt

Microphone capture uses the default input device. Allow microphone access in your operating system, some platforms also require PortAudio to be installed separately.

Data

The organizers supplied X_train.npy (14721 x 417), X_test.npy (22568 x 417), and y_train.npy (sample ID and label). Feature column 0 is the sample ID; the remaining 416 values are 32 time steps x 13 MFCC coefficients. Label -1 means other; labels 0 through 9 are digits. Padding (-9999999) becomes zero.

Training uses these arrays directly. The supplied wav2mfcc.py and transform_wav.py generated the MFCC representation. Competition audio and datasets are not redistributed.

Train

Place the data arrays in the repository root, or use --data-dir.

python main.py train
python main.py train --full-data

Development training uses a stratified 90/10 split and fits the scaler only on training samples. Full-data training uses all labelled samples without validation. Both save the model and fitted scaler to checkpoints/model.pt. Use --checkpoint to choose another output path.

The matrix below represents the stratified development validation split, cells show sample counts and row percentages.

Predict a WAV

python main.py predict path/to/audio.wav

Prediction prints a [0-9] digit or other. Use --checkpoint to select another model. WAV preprocessing uses the organizers' MFCC representation and boundary-padding behavior. Short inputs are padded and inputs over 32 MFCC frames are truncated.

Predict from the microphone

python main.py record

After the countdown and Speak!, capture lasts 1 second. Each run overwrites:

• recordings/latest.wav
• recordings/latest_spectrogram.png
• recordings/latest_prediction.png

These let you listen to the recording, inspect its spectrogram, and compare model scores across all 11 classes.

Structure

• main.py: commands, training, evaluation, and checkpoints.
• model.py, dataset.py, preprocessing.py: network and array preparation.
• audio_preprocessing.py, inference.py: WAV prediction.
• recording.py, plots.py: microphone capture and plots.
• docs/images/: selected documentation figures.

Generated recordings/, runtime_outputs/, and checkpoints/ folders are ignored by Git.

Author

@Bortrex