Request access to espeech_podcasts_chunked

Access is gated because this repository contains speech data. Request access only for approved training, evaluation, or research use.

By requesting access, you agree not to redistribute files, attempt speaker re-identification, or use the data for unauthorized voice impersonation.

Log in or Sign Up to review the conditions and access this dataset content.

espeech_podcasts_chunked

This is a gated Russian chunked speech dataset from instinct-org. It contains podcast speech segments and transcript text for speech-to-text training, evaluation, and data preparation workflows.

Current Audio Availability

This repository currently has a mixed audio-byte state:

  • data/train-00000-of-00130.parquet through data/train-00010-of-00130.parquet contain embedded audio bytes.
  • data/train-00011-of-00130.parquet through data/train-00129-of-00130.parquet preserve segment metadata and transcript text, but the audio.bytes field is not embedded. Those rows reference the original local source paths and are not directly usable for remote training or forced alignment without restoring the source audio.

For STT forced alignment or training that requires byte-backed audio rows, use instinct-org/espeech_podcasts_chunked_speech_restorised as the repaired source candidate until the original FLAC tree is available and this base repository can be reconstructed exactly.

Language

  • Primary language: ru (Russian)

Intended Use

  • speech-to-text training and evaluation after access approval
  • internal dataset curation, quality checks, and model evaluation
  • transcript and segment-metadata inspection

Data Notes

  • Columns include audio metadata, transcript text, speaker label, podcast id, segment index, and segment-level timing fields.
  • Existing chunk-level timing fields should be treated as segment metadata.
  • For word, character, token, or refreshed segment timestamps, rerun forced alignment from byte-backed audio and transcript text.

Access

This repository is gated with manual approval. Access requests are reviewed by the dataset maintainers. Users requesting access must provide their affiliation, intended use, and agreement to the dataset terms.

Collections

Safety And Rights

  • Do not redistribute the files outside approved access.
  • Do not attempt speaker re-identification.
  • Do not use the data for unauthorized voice impersonation.
  • Follow the source licenses and speaker-rights restrictions.
Downloads last month
507