ASR Code-Switch: 1,200 Code-Switching Utterances for Speech Recognition

1,200 code-switching utterances form a curated benchmark for evaluating commercial Automatic Speech Recognition systems. The dataset, created by Perle-ai, includes 300 samples each for four language pairs, such as Egyptian Arabic–English. It was last updated on May 21, 2026.

Use Cases

• Benchmarking ASR system performance on intra-sentential language switching based on the described utterances.
• Training or fine-tuning speech models for multilingual environments based on the code-switching samples.
• Analyzing error patterns in commercial ASR systems for specific language pairs like Egyptian Arabic–English.
• Researching linguistic phenomena in code-switched speech based on the curated utterance collection.

Strengths

• Contains 1,200 specifically curated code-switching utterances.
• Provides a structured benchmark with 300 samples per language pair.
• Focuses on intra-sentential language switching, a challenging scenario for ASR.

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

Perle-ai via Hugging Face.

Collection Method

Curated benchmark, likely gathered for research purposes as indicated by the linked arXiv paper.

Freshness

Last updated 2026-05-21 03:20:29; freshness should be verified.