Datasets:
audio audioduration (s) 3.01 28.6 | label class label 5
classes |
|---|---|
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both | |
0mix_both |
Libri2Mix, 16 kHz, dev and test
The dev and test splits of Libri2Mix: two LibriSpeech speakers per mixture, with and without WHAM! noise. These are the files used for the Libri2Mix experiments of the agentic-ASR project (Whisper with an LLM router that decides when to call speech separation). There is no train split.
How it was made
Generated with the official LibriMix script, scripts/create_librimix_from_metadata.py from
JorisCos/LibriMix at commit b898ca3, using the official metadata
(libri2mix_dev-clean.csv, libri2mix_test-clean.csv):
create_librimix_from_metadata.py --n_src 2 --freqs 16k --modes max min --types mix_clean mix_both
Inputs: LibriSpeech dev-clean and test-clean (OpenSLR 12), WHAM! noise cv and tt.
Layout
{max,min}/
{dev,test}/
mix_clean/ s1 + s2
mix_both/ s1 + s2 + noise
s1/ speaker 1 source
s2/ speaker 2 source
noise/ WHAM! noise
metadata/ LibriMix mixture CSVs (paths, lengths)
transcripts/
{dev,test}.csv ground-truth transcripts, one row per mixture
max: the mixture is as long as the longer utterance; the shorter one is zero-padded at the end.min: the mixture is cut to the shorter utterance.devcomes from LibriSpeech dev-clean,testfrom test-clean.- All audio is 16 kHz mono WAV. Paths inside
metadata/*.csvpoint to the machine where the data was generated; use the file names, not the absolute paths.
Transcripts
LibriMix ships no text, so transcripts/{dev,test}.csv gives the ground truth for each mixture, taken from the
LibriSpeech *.trans.txt files (dev-clean for dev, test-clean for test). Columns:
mixture_ID, s1_utt_id, s1_speaker, s1_text, s2_utt_id, s2_speaker, s2_text
mixture_IDis<s1_utt_id>_<s2_utt_id>, the same name as the WAV files, sos1/<mixture_ID>.wavis the speech ofs1_textands2/<mixture_ID>.wavthat ofs2_text.maxandminuse the same mixtures, so one file per split covers both. Inminmode the longer utterance is cut to the length of the shorter one, so its transcript is longer than what is audible; usemaxfor ASR.- The text is LibriSpeech's original: upper case, no punctuation.
| split | file | mixtures |
|---|---|---|
| dev | transcripts/dev.csv | 3000 |
| test | transcripts/test.csv | 3000 |
File counts (counted when the data was uploaded):
| mode | split | folder | files |
|---|---|---|---|
| max | dev | mix_both | 3000 |
| max | dev | mix_clean | 3000 |
| max | dev | noise | 3000 |
| max | dev | s1 | 3000 |
| max | dev | s2 | 3000 |
| max | metadata | csv | 8 |
| max | test | mix_both | 3000 |
| max | test | mix_clean | 3000 |
| max | test | noise | 3000 |
| max | test | s1 | 3000 |
| max | test | s2 | 3000 |
| min | dev | mix_both | 3000 |
| min | dev | mix_clean | 3000 |
| min | dev | noise | 3000 |
| min | dev | s1 | 3000 |
| min | dev | s2 | 3000 |
| min | metadata | csv | 8 |
| min | test | mix_both | 3000 |
| min | test | mix_clean | 3000 |
| min | test | noise | 3000 |
| min | test | s1 | 3000 |
| min | test | s2 | 3000 |
Total: 60016 files.
License and attribution
The mixtures contain WHAM! noise, so this dataset is released under CC BY-NC 4.0 (non-commercial use only).
- LibriSpeech: V. Panayotov, G. Chen, D. Povey, S. Khudanpur, "LibriSpeech: an ASR corpus based on public domain audio books", ICASSP 2015. CC BY 4.0.
- WHAM!: G. Wichern et al., "WHAM!: Extending Speech Separation to Noisy Environments", Interspeech 2019. CC BY-NC 4.0.
- LibriMix: J. Cosentino, M. Pariente, S. Cornell, A. Deleforge, E. Vincent, "LibriMix: An Open-Source Dataset for Generalizable Speech Separation", arXiv:2005.11262, 2020.
If you use this data, please cite the three works above.
- Downloads last month
- 153