Datasets:

Zineb
/

SonicCaps

audio_id stringlengths 11 11	caption_text listlengths 1 1
004NnY1farU	[ "Wind is blowing hard and waves are crashing on the shore." ]
0Om5Qjpllc4	[ "A male speaks while typing on a keyboard and clicking a mouse." ]
BD_qj9NSmUo	[ "A mid-frequency engine idles steadily with no other ambient sounds present." ]
1Cwdadv0Pbo	[ "A high-revving engine accelerates rapidly, its pitch rising sharply." ]
7lyJPIRPFHc	[ "A woman speaking clearly, her voice steady and articulate, shares her experience from a stage, the ambient sound of a quiet audience listening in the background." ]
hxU3NEJ6DW8	[ "A man speaks, followed by fast typing on a keyboard." ]
ih6rX2U59MI	[ "A person saws through wood, the blade scraping rhythmically against the grain." ]
yzEIYjxrbh4	[ "A baby laughs and babbles while adults chuckle in the background." ]
5e8sX3AyuDs	[ "Wind blowing softly as a sheep bleats twice in the distance." ]
nsJTiaEnciU	[ "Ducks quacking persistently as vehicle traffic hums in the background, then an adult male speaks briefly." ]
yglEqqemgFQ	[ "Motorboat engine running, water gurgling, and plastic thumping followed by metal creaking." ]
PXBvcucFV3A	[ "Humming of a large vehicle passing, followed by a hiss." ]
-3gSkrDKNSA	[ "A high-performance engine roars and fades into the distance." ]
V_aZ9aphY0U	[ "A high-pitched child speaks, followed by a loud pop and a brief echo." ]
B9VhcnX7kB4	[ "An engine idles while people talk and laugh nearby." ]
4m3Rqh1Mt1o	[ "A woman sings in a high-pitched voice while children laugh and talk excitedly." ]
4chQUsN0_YM	[ "A woman speaks in a formal tone, her voice clear and steady, with faint ambient room tone in the background." ]
kGPbA9wowRk	[ "Traffic moves with distant horns and a warning siren wailing." ]
d0-iUOthFcU	[ "A goat bleats as a woman and then a man speak in English." ]
7W0kSinVMxo	[ "A man speaks to a laughing crowd as they clap and murmur." ]
n_1blRUJebs	[ "A person snores at moderate volume nearby several times." ]
7BilNP2ejF8	[ "Water is dribbling down while music is playing over it." ]
7GtffbQkQHY	[ "Revving car engine idling and accelerating repeatedly." ]
2JgoGMg1v5Q	[ "Emergency sirens wail loudly in the night." ]
Gw1ioIGn9A4	[ "A gunshot echoes followed by a man screaming and speaking in pain." ]
3DOIkBRXnt4	[ "A cat meows and hisses while a man speaks to it." ]
3rCeT4gjGIk	[ "A man speaks in a language other than English, followed by loud applause from a crowd." ]
9ggN17UWxyU	[ "A woman speaks while a sewing machine operates in the background." ]
fLj8wKWa9k4	[ "Rain falls as thunder rumbles and a train horn sounds." ]
2aEhQ7vV6K0	[ "A female burps, laughs, and speaks in a casual indoor setting." ]
6A8zggJBXXw	[ "A woman speaks emotionally and smiles, prompting laughter from the audience in a large room." ]
BIhigCnEZOk	[ "High-pitched buzzing abruptly stops, followed by a loud thump and a click." ]
Bsx6eoI3NuQ	[ "Humming of an oncoming train with a honking horn and high-power whooshing as it passes." ]
GplpxGNK058	[ "Ducks quack as a man speaks, then a child yells in the background." ]
ISM2HiR85eY	[ "Humming and rustling accompany a man's calm greeting, abruptly interrupted by another man's loud, angry shouting." ]
fuX2vvOMFFc	[ "Pressurized air sprays continuously while a low motor rumble persists in the background." ]
BMRDQvCMIP8	[ "A machine emits repetitive beeps, followed by keyboard typing and a long, low bleep, all over a quiet hum." ]
242HiZmS644	[ "A motor engine sputters to life and settles into a steady idle." ]
1wpiL0AzEmM	[ "Gurgling and splashing water from a faucet in a bathroom." ]
2K139PoxG4Q	[ "A female voice speaks passionately, followed by a male voice speaking calmly, both in English." ]
0iWWVbtR-Rs	[ "A car engine cranks repeatedly before starting, with a hiss of air released." ]
8nd-jH9bfqk	[ "A power tool drills loudly, then stops abruptly as a male voice speaks in the background." ]
1bGr2lWrQRo	[ "Sheep bleat and grunt in a field." ]
mCNrz4nTKpQ	[ "Applause and cheering fill the air as a crowd expresses enthusiastic approval." ]
-pgmwA5EGbM	[ "A man speaks in a foreign language while flipping food in a pan, the sizzling sound of cooking filling the air." ]
9z8p5yPYblU	[ "A loud engine idles and vibrates nearby as a person speaks." ]
_WmmnVDID70	[ "Wind gusts blow into the microphone as thunder claps in the distance." ]
MUPn4gr56HY	[ "A baby cries while a woman speaks gently in a quiet indoor setting." ]
30Sm6NPl_rI	[ "Multiple dogs bark loudly as people talk nearby." ]
AuZ8Ae-4q5g	[ "Wind blowing, water splashing, and a motor running accompany a man speaking." ]
-C-jmSf9mTc	[ "Applause and cheering fill the air as a crowd reacts to a speaker." ]
1hL4MSPzMZ0	[ "Rapping is interrupted by a woman speaking in English." ]
1PyOibWAHL8	[ "Birds flap their wings and chirp in the background." ]
VSs-IsYZQGA	[ "Ocean waves crash loudly while wind blows hard into the microphone, as a man and a young boy speak in English." ]
3q7vSDImCE0	[ "Using a sewing machine, a man speaks in Spanish." ]
0bk8T9DGK18	[ "An adult female speaks in Spanish, followed by a dog whining." ]
a-cyYxQtiKo	[ "A bus engine runs with hissing air brakes, followed by a loud rush of wind." ]
3Y9pN7dCHik	[ "Beeping tones are followed by an adult male speaking in English." ]
esKgI564dPc	[ "A man speaks over the sound of wind blowing." ]
WC0IU37nI3U	[ "Birds chirp and coo, a dog barks, and a person coughs in the background." ]
EUIfNA0-bn8	[ "A woman speaks clearly and calmly in English, with no other discernible sounds in the background." ]
9OEoRX_f5Rw	[ "An aircraft engine roars to life with a deep, intense vibration." ]
37Upn7rqtvY	[ "Rain falls while a male voice speaks in German." ]
6LmUCyxN-Hg	[ "A man speaking clearly over the idle of a large vehicle, with speech in English." ]
1M9j-Cv5gLk	[ "Spraying sound is repeated in the background." ]
gWTZtnkXWpA	[ "Revving engine followed by squeaking brakes and loud metallic banging." ]
cvG5woVZtB4	[ "Rooster crows, birds chirp, and wind blows in the background." ]
f9XZIK-w0l4	[ "A man speaks while a baby nearby imitates speech sounds with high-pitched vocalizations." ]
71L97wy61yo	[ "A person snores rhythmically, with faint indistinct speech in the background." ]
B7EBdtqrllY	[ "A bus engine accelerates loudly, growing louder as it approaches and passes by." ]
BulIh3C-oY8	[ "A man speaks in a reverberant space, his voice echoing slightly as a crowd murmurs in the background." ]
Jnxo_obbopU	[ "Rain falls while a person speaks in Dutch." ]
566xlVKWwd4	[ "Typing sounds on a keyboard with a person breathing nearby and faint voices in the background." ]
6H3oX7XeKXs	[ "A person snores softly, then abruptly starts chuckling and laughing." ]
8RFjpTF_U7c	[ "Rapid gunfire echoes, followed by laughter and indistinct male speech." ]
vq9-fuuWV3k	[ "Two men having a conversation in Italian, one speaking at a fast pace." ]
-q7jVLJGHFE	[ "A large motor starts and runs loudly in an otherwise quiet environment." ]
cTnIjeWtYNc	[ "Water is poured into a container, followed by a child speaking in Mandarin Chinese." ]
8FNag4gPih8	[ "A woman speaking with a motorcycle engine revving in the background." ]
2-bnHoVlplo	[ "Crinkling followed by a woman speaking, then laughing in English." ]
FW6okWDWqG8	[ "Steam hisses nearby as people talk and a vehicle drives by." ]
5rGmGTCOsLI	[ "An engine idles roughly, accompanied by loud, intermittent squeaking." ]
1fE1-5KgCr0	[ "Quiet dog panting and shuffling sounds accompany a man's muffled laughter." ]
KEpp8-gamNo	[ "Horse clip-clop sounds are heard as an engine runs nearby and two people speak." ]
--N8lbFywRg	[ "Water trickles and splashes gently, creating a soothing ambient sound." ]
HJefzDdfP7Q	[ "A baby cries loudly while children talk in the background." ]
BCfvalVyB_A	[ "Speedboat engine revving, water splashing as it slows down." ]
487F6Pd2xAc	[ "Speaking, a woman discusses equipment while a person coughs in the background." ]
zuea6r_jasg	[ "Wind gusts and crashing ocean waves accompany a steady engine hum." ]
KxOF078vkFc	[ "Horse hoofs clop on pavement as wind blows, ending with a door slamming shut." ]
-YfsOuwDYh0	[ "A man closes a car door, speaks, and demonstrates a car locking mechanism." ]
s6eNfjjlzWw	[ "Pressurized air hisses as a large engine idles, a man's voice speaking with childlike wonder." ]
7cZdh62FjUU	[ "Water splashes and squeaks while a man speaks in English." ]
4O9eXzlv_Fc	[ "A door opens with a hiss, then closes with a soft thud in a quiet room." ]
E-RlDn9H2Yg	[ "A man and woman speak, a baby cries, and steam hisses." ]
ST0BAScAtgc	[ "A man speaks, a door slams, fire crackles, compressed air sprays, and a man laughs." ]
vyGmFKN8HxE	[ "A large bell tolls rhythmically, echoing in the air." ]
4-FKeE3M2oc	[ "An engine idles, then revs loudly and quickly to a high speed." ]
Yp639DmB6Sc	[ "Gurgling water accompanies a man speaking Japanese." ]
0iEi3WkLV_g	[ "Pigeons coo and flap their wings as a man calls out in the distance." ]

End of preview. Expand in Data Studio

SonicCaps

SonicCaps is a large-scale audio captioning dataset comprising ~15M captions paired with ~700k audio clips, generated with a multi-modal LLM (Qwen3-Omni) conditioned jointly on audio and text. Unlike prior datasets that provide a single caption per clip, SonicCaps contains ~24 captions per audio generated through structured prompt engineering and few-shot prompting, explicitly reflecting the one-to-many nature of auditory perception. These captions span multiple styles and levels of granularity, yielding both fidelity-focused and diversity-focused captions. Our fidelity-focused captions better align with human subjective judgments of caption quality, while the caption diversity of SonicCaps consistently improves audio–text retrieval performance compared to our baselines. We release two specialized models, SonicCLAP_AR for audio retrieval and SonicCLAP_MOS for perceptual quality assessment.

Paper: SonicCaps: Large-Scale Diverse and Fine-Grained Captioning for Improved Audio-Retrieval
Audio sources: Freesound, AudioCaps, BBC Sound Effects, AudioSet Strongly-Labeled
Models: SonicCLAP_AR and SonicCLAP_MOS will be released shortly.

Dataset Description

Each audio clip is associated with four caption types designed to capture different levels of semantic granularity.

Caption subset	Description	Avg. #/audio	Avg. length	Vocabulary	% Unique Captions
`main`	Fidelity-focused, factual description	1	11.2 words	26K	92%
`rephrased`	Diverse linguistic rephrasings	10	9.1 words	41K	87%
`rephrased-short`	Short, query-like captions	10	4.5 words	47K	52%
`tags`	Semantic keyword tags	3	3.3 words	18K	51%

Audio waveforms are not redistributed. Only captions and their corresponding audio identifiers are released.

Please refer to the license terms of each original source (linked below) to obtain the audio.

Dataset Composition

Audio source	# Audios	# Captions (SonicCaps)	Audio download link
Freesound	515K	~12.3M	https://huggingface.co/datasets/Meranti/CLAP_freesound
BBC Sound Effects	31K	~0.7M	https://huggingface.co/datasets/cvssp/WavCaps/tree/main/Zip_files/BBC_Sound_Effects
AudioSet Strongly-Labeled	108K	~2.6M	https://huggingface.co/datasets/cvssp/WavCaps/tree/main/Zip_files/AudioSet_SL
AudioSet (AudioCaps subset)	49K	~1.2M	https://huggingface.co/datasets/OpenSound/AudioCaps

Download

These snippets only demonstrate how to identify and link SonicCaps captions to their original audio sources. They load metadata only (IDs, filenames, timestamps), not the audio itself, which must be downloaded separately from the linked source repositories.

from datasets import load_dataset, hf_hub_download
import pandas as pd
import json

caption_type = "main" # rephrased, rephrased_short, tags

# AudioCaps
soniccaps = load_dataset("Zineb/SonicCaps", caption_type, split="AudioCaps").to_pandas()
audio_meta = load_dataset("d0rj/audiocaps", split="train").to_pandas()

merged = audio_meta.merge(soniccaps, left_on="youtube_id", right_on="audio_id", how="inner").rename(columns={"caption_text": f"soniccaps_{caption_type}"})

# FreeSound
soniccaps = load_dataset("Zineb/SonicCaps", caption_type, split="FreeSound").to_pandas()
csv_path = hf_hub_download(repo_id="Meranti/CLAP_freesound", repo_type="dataset", filename="freesound_meta.csv")
audio_meta = pd.read_csv(csv_path)

merged = audio_meta.merge(soniccaps, left_on="audio_filename", right_on="audio_id", how="inner").rename(columns={"caption_text": f"soniccaps_{caption_type}"})

# BBC_Sound_Effects
soniccaps = load_dataset("Zineb/SonicCaps", caption_type, split="BBC_SoundEffects").to_pandas()
json_path = hf_hub_download(repo_id="cvssp/WavCaps", repo_type="dataset", filename="json_files/BBC_Sound_Effects/bbc_final.json")
audio_meta = pd.DataFrame(json.load(open(json_path))["data"])

merged = audio_meta.merge(soniccaps, left_on="id", right_on="audio_id", how="inner").rename(columns={"caption_text": f"soniccaps_{caption_type}"})

# AudioSet Strongly-Labeled
soniccaps = load_dataset("Zineb/SonicCaps", caption_type, split="AudioSet_SL").to_pandas()

json_path = hf_hub_download(repo_id="cvssp/WavCaps", repo_type="dataset", filename="json_files/AudioSet_SL/as_final.json")
audio_meta = pd.DataFrame(json.load(open(json_path))["data"])
merged = audio_meta.merge(soniccaps, left_on="id", right_on="audio_id", how="inner").rename(columns={"caption_text": f"soniccaps_{caption_type}"})

Some entries (~1k) from the original audio datasets may not have a corresponding caption in SonicCaps, expect a small fraction of unmatched rows (how="inner" drops them automatically).
FreeSound, BBC_Sound_Effects and AudioSet_SL audio require downloading .tar/.zip archives separately (see linked source repos)

Models

We release two CLAP models trained on SonicCaps with a multi-caption sampling strategy:

SonicCLAP_AR : optimized for audio-text retrieval (best R@k on AudioCaps-Val and commercial benchmarks)
SonicCLAP_MOS : trained on main captions, best correlation with human perceptual judgments (MOS)

Code, training configs, and inference scripts for both models are available on GitHub (linked above).

License

SonicCaps captions are released under CC-BY-NC-4.0. Audio must be obtained from the original sources, each governed by its own license: Freesound, BBC Sound Effects, AudioSet.

Citation

@inproceedings{XXXX,
  title={SonicCaps: Large-Scale Diverse and Fine-Grained Captioning for Improved Audio-Retrieval},
  author={Zineb Lahrichi, Marc Ferras, and Gaël Richard and Geoffroy Peeters.},
  year={2026}
}

Acknowledgment

We thank Pro Sound Effects for granting permission to release captions associated with their datasets.

Downloads last month: -

Paper for Zineb/SonicCaps

SonicCaps: Large-Scale Diverse and Fine-Grained Captioning for Improved Audio-Retrieval

Paper • 2609.02343 • Published 2 days ago