AudioDataset

Dataset Overview

Speech Datasets

1. VCTK v0.92

   • Contains 110 English speakers with various accents, each reading approximately 400 sentences.
   • All recordings converted to 48 kHz.
   • Training set size: To be determined.

2. VoiceBank-DEMAND

   • Single-channel human speech, 48 kHz.
   • Speech with noise, where noise is sourced from the DEMAND dataset.
   • 28‑speaker version:
     • Training set includes 11,572 utterances (≈9.4 hours).
     • Test set includes 824 utterances (≈0.6 hours).
     • Validation set (separated from training) includes 770 utterances (≈0.6 hours), remaining training set includes 10,802 utterances (≈8.8 hours).
   • 56‑speaker version: To be determined.

3. LibriSpeech

   • Approximately 1000 hours of 16kHz read English speech corpus.
   • Training set: 100h + 360h + 500h (total 960h).
   • Development and test set details: To be determined.

4. DNS-Challenge 5

   • Contains multilingual clean speech and various noises.

Music Datasets

1. StarNet
   • Contains 104 classical music tracks at 48 kHz, sourced from corresponding free MIDI files.
   • Track details:
     • xxx.0.wav: Clarinet–vibraslap mix
     • xxx.1.wav: Clarinet track
     • xxx.2.wav: Vibraslap track
     • xxx.3.wav: String–piano mix
     • xxx.4.wav: String track
     • xxx.5.wav: Piano track

Audio Mixture Datasets

1. Libri2Mix

   • Details: To be determined.

2. Divide and Remaster (DnR)

   • Single-channel mixed audio, containing speech, music, and sound effects/background tracks, 44.1 kHz.
   • Audio sources: LibriSpeech (speech), Free Music Archive (music), FSD50k (sound effects).
   • Training set includes 3,406 mixes (≈57 hours).
   • Validation set includes 487 mixes (≈8 hours).
   • Test set includes 973 mixes (≈16 hours).

3. MUSAN

   • Corpus containing music, speech, and noise, 16 kHz, suitable for voice activity detection (VAD) and music/speech discrimination.
   • Speech: 426 recordings (≈60 hours), sourced from LibriVox and the US government.
   • Music: 660 recordings (≈42 hours), sourced from Jamendo, Free Music Archive, Incompetech, and HD Classical Music.
   • Noise: 930 recordings (≈6 hours), sourced from Free Sound and the Sound Bible.