A speech dataset can connect one recording to many training objectives, but most collections provide only a single transcript. This release takes a broader approach: it pairs speech with multiple aligned text targets, making it possible to train and evaluate systems for transcription, translation, and instruction-following without treating each task as an isolated dataset.
What Sets It Apart
- Six text target types include normalized transcripts, inverse text normalization, captions, code-switched text, speech question-answering, and translations. This lets researchers study output formats and tasks against the same source utterance.
- The public manifests cover 48 languages and 463,062 hours of source speech across 16 corpus entries; approximately 165,300 hours pass the pipeline’s quality gate. The scale is substantial, but the quality flag and target availability must be checked record by record.
- Translation targets run from non-English speech to English and from English speech to 31 languages. Translations are embedded in the records with quality scores, but a target key may contain an empty rejected candidate.
- The dataset provides labels and audio identifiers, not audio files. The source audio must be acquired separately under each corpus’s terms, and the released materials have subset-specific licensing.
Who It’s For
A good fit for researchers building multilingual ASR, speech translation, or speech-instruction models who need several aligned targets and can handle source-audio retrieval and license tracking. Look elsewhere if you need a turnkey audio download, uniformly complete translations, or human-verified labels: the labels are automatically generated, coverage varies, and the quality gate is not a guarantee that every target is usable.
