NADI2025 Subtask2: Multidialectal Arabic Speech Recognition Benchmark

NADI2025 Subtask2 is a benchmark for developing Automatic Speech Recognition systems that transcribe Arabic speech across multiple dialects. The dataset is hosted by UBC-NLP on Codabench and was last updated on June 12, 2025. It provides training and validation data for handling phonetic and dialectal variation, with a private test set to be released.

Use Cases

• Benchmarking ASR model performance on multidialectal Arabic speech based on the described phonetic and dialectal variation.
• Training zero-shot ASR models using the provided validation data.
• Fine-tuning ASR models on specific Arabic dialects using the provided training set.
• Developing techniques for handling phonetic variation in speech recognition systems.

Strengths

• Focuses on a specific, challenging task: multidialectal Arabic ASR.
• Provides distinct training and validation splits for model development.
• Managed by a known research entity (UBC-NLP) and hosted on a benchmarking platform (Codabench).

Limitations

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

Provenance

Source

UBC-NLP

Collection Method

Likely gathered for the NADI-2015 Subtask-2 competition.

Freshness

Last updated 2025-06-12 20:19:46; freshness should be verified.

Geography

Covers multiple Arabic dialects; specific geographic coverage is not detailed.