Hugging Face's logo

Datasets:
Kppwdfgu1
/
wordts_smoke

Tasks:
Text-to-Speech
Automatic Speech Recognition
Modalities:
Audio
Text
Formats:
soundfolder
Languages:
English
Yoruba
Igbo
Size:
1K - 10K
Tags:
audio
nigerian-languages
nigerian-english
nigerian-pidgin
forced-alignment
Libraries:
Datasets
Dataset card Data Studio Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This is an unreviewed speech-alignment dataset. Explain your intended use so the owner can manually review your request.

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

wordts_smoke

This dataset contains lossless FLAC chunks derived from 30 Nigerian-language, Nigerian English, and Nigerian Pidgin recordings. Access requests require manual approval by the repository owner.

Contents

  • Chunks: 2,217
  • Chunk audio duration: 3.673 hours
  • Source transcript rows represented: 4,328
  • Standalone audio-tag chunks: 0
  • Non-music tag rows attached to nearest same-speaker speech: 102
  • Music annotation boundaries: 227
  • Non-speech annotations: 553 (also retained in the audit manifest)
  • Immutable source rows: 4,778
  • Non-music tag occurrences: 331
  • Recordings: 30
  • Audio: source-rate, source-channel-count, PCM-16 FLAC
  • Timestamp source: transcript_sbpn
  • Quality status: unreviewed

Chunking policy

Consecutive transcript rows are merged only when they have the same recording-local speaker label, contain no intervening non-speech annotation, overlap or have a gap of at most 2.0 seconds, and produce a chunk no longer than 40.0 seconds. A single source row longer than the maximum is retained unsplit because no internal row-level cutting point exists.

Non-music tag-only rows such as <laugh> and <sigh> follow the versioned deterministic association rules: inline annotations stay in place; same-speaker bridges are placed between both speech rows; leading/terminal rows are prepended/appended; conflicts or gaps over 2.0 seconds remain unresolved. Every SBPN/finalized-diarization timestamp stays unchanged. Music annotations are removed from published text and act as hard chunk boundaries.

Important limitations

  • Speaker labels are anonymous and local to each recording. Speaker 1 does not identify the same person in different recordings.
  • The timestamps come from transcript_sbpn and can still contain alignment or boundary errors. These chunks are intended for manual review and further automated QA, not direct TTS training.
  • Music, singing, overlapping speech, noise, and transcript errors have not been filtered out.
  • The source recordings were compressed media; FLAC avoids another lossy generation but cannot restore information removed by the source codec.
  • No license is asserted here. Users must independently confirm rights, consent, and permitted uses before training or redistribution.

Fields

The Dataset Viewer reads train/metadata.jsonl. Important fields include file_name, text, speaker_local, recording_id, exact start/end times, chunk_type, audio-tag timing/context provenance, source-row indices, gap statistics, and the over_40s_unsplittable flag.

manifests/chunks.jsonl contains the same records in JSONL form. manifests/immutable_event_ledger.jsonl preserves every raw source row and manifests/derived_tag_events.jsonl records reversible placement, confidence, context, normalization, and review fields. manifests/source_summary.csv and manifests/non_speech_events.csv provide recording-level and event-level audit information.
Downloads last month
11
Total file size:
682 MB