README.md

metadata

pretty_name: Bagpiper Pretraining Data Collection
language:
  - en
task_categories:
  - automatic-speech-recognition
  - text-to-audio
tags:
  - audio
  - speech
  - music
  - sound
  - multimodal
  - rich-caption
  - opus
  - speech-language-model
size_categories:
  - 100M<n<1B
viewer: false

Bagpiper Pretraining Data Collection

Opus-compressed audio with rich captions and curation subset labels, in the archive layout the ESPnet speechlm dataloader reads directly.

This is the compressed counterpart to espnet/Bagpiper_PreTrain_Data. That repository ships HuggingFace-style Parquet with embedded audio bytes, which is convenient for load_dataset but is not loadable by the training code. This repository ships the archive format the loader expects, so it can be trained on after one preparation command.

Contents

25 source datasets, 392,417,320 utterances, 933,497 audio hours, 14.00 TB.

dataset	utterances	hours	GB	captioned
owsm_v4	71,573,818	309,024	4,535.5	100.0%
laion_audio_300m_part3	65,357,718	127,417	1,914.7	100.0%
laion_audio_300m_part2	65,357,718	126,350	1,901.4	100.0%
laion_audio_300m_part1	65,357,718	115,279	1,754.5	100.0%
laion_audio_300m_part4	65,357,715	98,189	1,528.1	100.0%
yodas_auto	24,907,459	67,254	1,008.2	99.2%
emilia_en	17,514,480	45,214	679.9	100.0%
yodas_manual	5,278,839	14,188	214.4	97.1%
laion_captioned_ai_music_snippets	3,400,994	8,144	128.5	100.0%
fma	2,972,622	8,185	126.0	100.0%
wavcaps	2,749,599	7,440	106.2	90.7%
audioset	2,053,857	5,705	84.9	100.0%
laion_in_the_wild_sound_events	474,546	928	13.8	100.0%
clotho_train	34,918	87	1.3	100.0%
clotho_development	3,837	24	0.3	100.0%
mmau_test_music	3,000	13	0.2	100.0%
mmau_test_speech	3,000	12	0.2	100.0%
clotho_aqa	4,880	12	0.2	100.0%
mmau_test_sound	3,000	10	0.1	100.0%
clotho_test	1,043	7	0.1	100.0%
librispeech_test_clean	2,620	5	0.1	100.0%
librispeech_test_other	2,939	5	0.1	100.0%
mmau_test_mini_music	334	1	0.0	100.0%
mmau_test_mini_speech	333	1	0.0	100.0%
mmau_test_mini_sound	333	1	0.0	100.0%
total	392,417,320	933,497	13,998.8	99.8%

Audio is Ogg/Opus, 16 kHz mono, about 33 kbps. Stereo sources are downmixed and everything is resampled to 16 kHz, which is what the model consumes. Opus accepts only 8, 12, 16, 24 and 48 kHz, so 22.05, 32 and 44.1 kHz sources were resampled rather than stored at their native rate.

Layout

audio/<dataset>/shard-NNNNNN.bin        Ogg/Opus blobs, concatenated
audio/<dataset>/shard-NNNNNN.parquet    per-shard index
audio/<dataset>/metadata.parquet        merged index for the dataset
subsets/<subset>/<dataset>.txt          utterance ids per curation subset
subsets/counts.json                     subset sizes
tools/prepare_local.py                  post-download preparation

prepare_local.py adds the following on your machine; they are not shipped, because they hold absolute paths:

rich_caption/<dataset>/dump/**          zstd caption archive (arkive_text)
data_jsons/<dataset>/{caption,<subset>}.json  loader manifests: all rows, and one per subset
registry.yaml                           ESPNET_DATASET_REGISTRY entries

Each row of metadata.parquet carries the loader columns utt_id, path, start_byte_offset, file_size_bytes, start_time, end_time, plus doc_id, dataset, sample_rate, channels, length, format, source_sample_rate, source_file_size_bytes, the caption as a plain string in rich_caption, and the curation labels subsets, category, direction.

Utterances that are time windows over a longer recording share one Opus blob and are distinguished by start_time and end_time, exactly as in the FLAC source.

Curation subsets

subsets is a list, because an utterance can belong to none, one or two of them. Membership comes from the stage-5 curation outputs. An utterance is only ever in the und and/or gen set of a single modality, so category and direction are a lossless flattening of the six base subsets. Nine further score-ranked subsets from the SFT selection are included where they exist.

subset	utterances	datasets
speech_und	85,023,000	15
speech_gen	105,595,360	15
sound_und	23,980,555	15
sound_gen	24,001,618	15
music_und	14,392,994	15
music_gen	14,073,000	15
speech_und_sft	49,211	12
speech_gen_sft	199,847	11
speech_gen_sft2	1,992,076	13
sound_und_sft	46,976	15
sound_gen_sft	169,467	12
sound_gen_sft2	1,840,892	15
music_und_sft	28,941	13
music_gen_sft	61,367	9
music_gen_sft2	957,323	15

Use

huggingface-cli download --repo-type dataset     espnet/Bagpiper_PreTrain_Data_Collection --local-dir ./bagpiper
python ./bagpiper/tools/prepare_local.py --root ./bagpiper
export ESPNET_DATASET_REGISTRY=./bagpiper/registry.yaml

The preparation step is required, and is not optional packaging. The archive stores absolute byte offsets and the loader uses the path column verbatim with no root override, so the metadata has to be pointed at wherever the files landed. The same command also builds the zstd caption archive the arkive_text reader expects, and writes one manifest per subset. It is idempotent.

Reading a clip without ESPnet needs only a seek and a decode:

import io, soundfile as sf, pyarrow.parquet as pq
# after prepare_local.py, so that `path` points at your copy
row = pq.read_table("audio/clotho_test/metadata.parquet").slice(0, 1).to_pylist()[0]
with open(row["path"], "rb") as f:
    f.seek(row["start_byte_offset"])
    wav, sr = sf.read(io.BytesIO(f.read(row["file_size_bytes"])))

This needs a libsndfile with Opus support, which means 1.0.29 or newer; 1.2.x reports the stored 16 kHz rate correctly.

Caveats

• Captions are machine generated and were not verified by a human. Rows whose generation hit the length limit are kept and flagged in caption_finish_reason.
• Caption coverage is not complete for every dataset; the table above gives the per-dataset figure.
• Re-encoding from FLAC to Opus is lossy and not reversible.
• No blanket license is declared. The underlying corpora carry their own terms and you are responsible for reviewing them.

Citation

@inproceedings{tian2026bagpiper,
  title     = {Bagpiper},
  author    = {Tian, Jinchuan and others},
  booktitle = {COLM},
  year      = {2026}
}