---
pretty_name: gdrive-transcript1-audio-chunks-20260805-source-06
task_categories:
- automatic-speech-recognition
- text-to-speech
language:
- yo
tags:
- audio
- forced-alignment
- diarization
- quality-filtering
gated: true
extra_gated_prompt: "Explain your intended use. Every access request is reviewed manually."
extra_gated_fields:
  Intended use: text
  Organization or affiliation: text
  I understand that the audio, transcripts, timestamps, speaker labels, and automated quality labels may contain errors: checkbox
---

# gdrive-transcript1-audio-chunks-20260805-source-06

This gated dataset contains 2,117 MP3 speech and audio-event chunks from
50 recordings. Access requires manual approval by the
repository owner.

## Audio selection

- Every complete source recording was separated once with HTDemucs.
- Full vocal stems were stored remotely as 48 kHz mono 96k MP3;
  no per-chunk source separation was performed.
- Librosa SNR was calculated independently on aligned original and saved-vocals
  slices using maximum frame RMS and the 10th-percentile RMS floor.
- Chunks with finite `snr_after_librosa_db - snr_before_librosa_db >= 0` were
  cut from the saved full vocals MP3. Chunks below 0 dB were cut from the
  original recording. Non-finite comparisons conservatively use the original.
- Noise was labelled in metadata and did not by itself trigger denoising.

## Counts

- Chunks: 2,117
- Standalone audio-tag chunks: 0
- Speech chunks retaining inline audio tags: 50
- Non-music tag rows attached to nearest same-speaker speech:
  55
- Music annotation boundaries: 238
- Duration: 6.626 hours
- Chunks using Demucs vocals: 1,152
- Chunks using original audio: 965
- Non-finite SNR comparisons using original fallback:
  11
- Music/beat-labelled chunks (metadata only): 1
- Noise-labelled chunks: 17
- Transcript source: `transcript_sbpn_diarized`

Automated quality decisions are conservative estimates and should be audited
before irreversible filtering or model training.