MoDiCoL: Modular Diagnostic Continual Learning Dataset for ASR

MoDiCoL is a speech dataset designed to study the robustness of ASR models to different drift factors in a controlled, continual setting. It is constructed using a systematic factorial design that enables a rigorous evaluation of linguistic, speaker, and acoustic variation with clearly defined experimental runs. The dataset combines real-world and synthetic speech with a configuration-dependent augmentation pipeline, and was created by TPekarekRosin.

Use Cases

• Benchmarking ASR model robustness to linguistic drift based on the described factorial design
• Studying speaker variation effects on ASR performance in a controlled setting
• Evaluating acoustic drift adaptation strategies using the synthetic and real-world speech components
• Developing continual learning algorithms for speech recognition based on the dataset's structured experimental runs

Strengths

• Designed with a systematic factorial design for controlled evaluation of multiple drift factors
• Combines real-world and synthetic speech data to create varied test conditions
• Includes a configuration-dependent augmentation pipeline for generating specific experimental runs

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

huggingface

Collection Method

Constructed using a systematic factorial design combining real-world and synthetic speech with an augmentation pipeline.

Freshness

Last updated 2026-06-16 12:36:13; freshness should be verified