🎙️

SBPN: multilingual ASR models for Nigerian languages

• Speech recognition models for Yoruba, Hausa, Igbo, Nigerian Pidgin and Nigerian English, which outperform other open-source models on these languages.
• Two sizes, fine-tuned from NVIDIA Parakeet with NeMo: base (120M parameters) and large (600M parameters).
• Described in the paper Sometin Beta Pass Notin, accepted at SLT 2026.

• base model
• large model
• paper

🧩

Speech Language Models

• Research implementations of speech language models, focusing on recent advances in speech language modelling with acoustic codes.
• Includes VQ-VAE, wav2vec, vq-wav2vec and wav2vec 2.0, following the original papers closely.

• code

🔊

Iterative Vocoders for Phase Retrieval

• Implements the iterative phase-retrieval vocoders from Beyond Griffin-Lim (Peer et al.).
• Griffin-Lim, Fast Griffin-Lim, RAAR, Difference Map, ADMM, and hybrids that combine any of them.

• code

📊

ASR Statistical Significance Testing

• Tests whether the word error rate difference between two ASR models is statistically significant, by computing a confidence interval with bootstrap sampling.
• Supports blockwise bootstrap, e.g., sampling by speaker, gender or age.

• code

SAINT: Neural Networks for Tabular Data

• Unofficial PyTorch implementation of SAINT (row attention and contrastive pre-training), built in a week with a team of 8 research students.
• Works with any tabular dataset for binary or multiclass classification, with PyTorch Lightning and Hydra configs.

• code

Yoruba Speech Dataset: a dataset for low-resource speech tasks

• 3 hours of audio recordings from a native Yoruba speaker.
• Can be used for low-resource speech tasks.

• code

Exploring Transfer of Self-supervised English Pretrained Model to Yoruba and French Languages in Very Low-Resource Settings

• A self-supervised (CPC) model pretrained on a large amount of English speech is fine-tuned on 1 hour of labelled Yoruba and French audio.
• We compare this to training from scratch on 1 hour of audio in the target language. Audio samples are from a single speaker.
• Surprisingly, fine-tuning on a single speaker gave better results for Yoruba. Many factors could contribute, including speaker gender (male for Yoruba, female for French) and recording quality.

• code
• video demo

Exploring Detectron2 for Object Detection Tasks

• Instance segmentation, pose estimation and object tracking with the Detectron2 library.
• Pretrained models fine-tuned on custom tasks.

• code
• tracking demo

Cassava Disease Classification

• Classify cassava plants as healthy or as one of 4 disease classes.
• Experimented with various CNN architectures.
• Explored different ways of dealing with imbalanced datasets. I wrote a blog post on what worked for me.

• code

DNA Sequence Classification using Kernel Methods

• Predict whether a DNA sequence region is a binding site using kernel methods.
• Experimented with mismatch, Gaussian and weighted-degree kernels.

• code