FLEURS Code-Switching Evaluation Set

Dataset Summary

This dataset is a synthetic code-switching evaluation set built from the google/fleurs corpus.
Each sample is a single long-form audio sequence (minimum 5 minutes by default) composed by concatenating short utterances from multiple languages.

The goal is to provide a controlled benchmark for testing ASR robustness when language switches happen frequently inside one recording.

How The Dataset Was Curated

• Source data: google/fleurs Parquet files loaded per language/split (default split: test).
• Languages used: en, fr, es, de, ru, it, pt, nl.
• Per-sample language mix: a random subset of 2 to 8 languages.
• Coverage constraint: each selected language appears at least once in the sample.
• Construction rule: utterances are randomly sampled and concatenated until sample duration reaches at least 300 seconds (default).
• Audio normalization: utterances are decoded to a common sampling rate (--target-sr, default 16kHz) and converted to mono when needed.

Columns

• id: unique sample index.
• audio: concatenated waveform and sampling rate.
• transcription: plain concatenation of chunk transcripts.
• transcription_tagged: transcript with inline language and timing markers for each chunk, formatted as
  <lang><start:SS.ss>text<end:SS.ss>.
• duration_sec: final sample duration in seconds.
• languages: languages selected for that sample.
• seed: per-sample random seed.

Limitations

• Code-switching is synthetic (concatenative), not natural conversational switching.
• Prosody, speaker continuity, and discourse-level transition cues are not preserved across joins.