tb-voice / README.md
Shadow0482's picture
Upload quality-filtered TB Voice 30-hour dataset
70c956b verified
|
Raw History Blame Contribute Delete
1.64 kB
metadata
language:
  - en
license: cc-by-4.0
task_categories:
  - audio-to-audio
tags:
  - speech
  - audio
  - full-duplex
  - english
  - conversational-speech
pretty_name: TB Voice - Full-Duplex English Speech

TB Voice

A curated English full-duplex conversational speech dataset.

Dataset Overview

Property Value
Total audio duration 29.98 hours
Target duration 30 hours
Selected recordings 114
Audio format FLAC
Sample rate 44.1 kHz (original sample rate preserved)
Channels 2
Language English
Conversation type Full-duplex
License CC BY 4.0

The selected recordings preserve both speaker channels and their original conversation timing.

Duration is calculated from the recording timeline. Two speaker channels count as one hour of conversation, not two.

Data Structure

tb-voice/
├── README.md
├── manifest.csv
├── filter_summary.json
└── data/
    └── train/
        ├── shard-000000.tar
        ├── shard-000001.tar
        └── ...

Each TAR shard contains paired files:

sample_key.flac
sample_key.json

The JSON files preserve the associated recording metadata.

Quality Filtering

Hierarchical filter settings applied:

  • Valid stereo audio.
  • Minimum recording duration of 10 seconds.
  • Minimum sample rate of 16000 Hz.
  • Minimum overall RMS level of -48.0 dBFS.
  • Speech activity measured using 20 ms frames.
  • Activity threshold of -45.0 dBFS.
  • Minimum activity of 2% across the recording.
  • Activity in both speaker channels.
  • Maximum clipping ratio of 0.5%.
  • Valid JSON metadata.