Multilingual Code-Switched Speech Data for 52 Languages

CS-FLEURS contains 300 hours of speech data across 113 unique code-switched language pairs. It includes both read and synthetic speech for developing and evaluating speech recognition and translation systems.

Use Cases

• Train multilingual speech recognition models on 52 languages using the provided audio and text pairs.
• Develop code-switched speech translation systems for the 113 unique language pair combinations.
• Evaluate model performance on the Read-Test subset containing 14 X-English pairs of read speech.
• Fine-tune text-to-speech models using the XTTS-Train subset of synthetic speech data.

Strengths

• Covers 52 languages, enabling multilingual model development.
• Includes 113 unique code-switched language pairs for specialized research.
• Contains 300 hours of speech data, providing substantial audio material.

Limitations

• Dataset size and specific row counts for audio files are not provided.
• The distribution of data across the 113 language pairs is unspecified, which may indicate imbalance.
• The proportion of read versus synthetic speech within the 300 hours is not detailed.

Provenance

Source

Hugging Face dataset authored by 'byan'.

Collection Method

Dataset includes both read speech and synthetic (XTTS) speech data.

Freshness

Last updated on 2025-09-17.