The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

Soundscapes Temporal Events

This dataset adds detailed, time-aligned annotations of speech, sound effects, and music to audio from LAION Soundscapes. The release is being uploaded. The planned first release contains 516,500 validated clips. The available file list shows which shards are currently ready.

The annotations are automatically generated. Event boundaries, transcripts, categories and descriptions can contain errors. Age, gender, delivery and emotion fields describe uncertain audible impressions. They are not verified personal attributes or a measure of a speaker's actual feelings.

Listen to examples and explore their timelines.

Data format

Audio and annotation records are grouped by key in WebDataset TAR shards under data/. Each record contains an original source audio file and a json document with an events array and source identifiers. Event times are in seconds from the start of the audio. Events may overlap. Categories are speech, sfx, and music. The annotation includes captions, and speech events may also contain speaker IDs, verbatim transcripts, transcript status, language, delivery, emotion, and perceived age and gender. Fields not applicable to an event are empty.

The source identifier refers to the original Soundscapes sample and source shard. Retain attribution and available upstream source details when reusing data. Original audio encoding varies by source.

# pip install huggingface_hub webdataset
from huggingface_hub import list_repo_files, hf_hub_url
import webdataset as wds

repo = "laion/soundscapes-temporal-events"
files = sorted(p for p in list_repo_files(repo, repo_type="dataset")
               if p.startswith("data/") and p.endswith(".tar"))
urls = [hf_hub_url(repo, p, repo_type="dataset") for p in files]
dataset = wds.WebDataset(urls, shardshuffle=False).decode()
for sample in dataset:
    annotation = sample["json"]
    audio_member = next((k for k in ("mp3", "wav", "flac", "aac", "ogg", "m4a") if k in sample), None)
    print(sample["__key__"], audio_member, annotation["events"][:1])
    break

The annotations describe audible events. Short sounds may have very short intervals and multiple sources may overlap in time. Speaker IDs are local to a clip. [inaudible] indicates unintelligible speech. Singing is annotated as music.

License and attribution

This release is provided under Creative Commons Attribution 4.0 International (CC BY 4.0) to the extent of the contributors' rights. Credit LAION e.V. and the underlying LAION Soundscapes dataset, and retain any applicable source attribution. As the source dataset notes, third-party rights in underlying sounds or other source material remain with their respective holders. Consult the source metadata for individual recordings.

Downloads last month
-

Space using laion/soundscapes-temporal-events 1