speechdatasamples/darija-6500-samples-20261026

Darija 6,500 sample

A private, curated sample of 6,500 audio segments with transcript text and parent/chunk identifiers.

Columns

• audio: source audio segment (FLAC)
• text: transcript
• parent_audio_id: source parent recording identifier
• chunk_order: segment order within the parent recording

Curation

The dataset combines 2,239 rows retained from the earlier reviewed sample with 4,261 newly selected rows. New rows were selected from CTC-scored data at nll_per_token <= 0.45, excluding every parent/chunk pair from the earlier 6,500-row candidate sample. Neighboring eligible chunks were grouped by parent recording.

For the new sample, one Gemma consensus transcript per parent recording was reviewed for Moroccan Darija. Parent groups whose example transcript indicated formal Arabic, another regional dialect, or a foreign language were excluded. This was a transcript-based review; the audio was not independently auditioned during this pass. Therefore dialect labels reflect conservative transcript review, not a formal language-identification audit.

Original source data was not deleted or modified. This repository contains a separate derived dataset.