Speaker turn labels and the reliablity of transcripts
Impressive scale and dataset effort!
Out of curiosity, do you have any plans to add speaker turn labels in future versions?
And since many transcripts come from YouTube’s auto-captions or open-source STT models—where word error rates (WER) can exceed 20% in noisy conditions, and perform even worse on named entities (brands, names, addresses) and low-resource languages. Do you have any plans or a roadmap to improve transcript quality?
When we built OleSpeech-IV from similar web streams, we added human-sourced speaker labels and word-level confidence scores so researchers can easily handle multi-speaker turns and filter noisy web captions out of the box: https://huggingface.co/datasets/olewave/OleSpeech-IV-2025-EN-AR-100
Great work expanding open multilingual speech resources for the community!