hamzahanif/memoni-audio

Memoni Audio Dataset

Cleaned, voice-activity-detected speech segments collected from online video sources for Arabic/regional dialect speech research.

Statistics

Metric	Value
Total segments	36,787
Total speech duration	5679.1 min (94.7 h)
Categories	15
Platforms	facebook, instagram, nan, tiktok, youtube
Channels	26

Processing Pipeline

1. Download — YouTube audio downloaded as MP3.
2. VAD — TenVad at 16 kHz mono; original sample rate preserved for all audio ops.
3. Chunking — Speech regions < 30 s kept as-is; >= 30 s split into ~30 s sub-chunks.
4. Denoising — dpdfnet (dpdfnet8_48khz_hr, attn_limit_db=12).
5. Export — 192 kbps MP3, peak-normalised to 0 dBFS.

Dataset Fields

Field	Type	Description
audio	Audio	MP3 file
unique_id	string	Segment identifier (e.g. 00042_seg0003)
source_video_id	string	Original video unique_id
youtube_link	string	Source YouTube URL
category	string	Content category
channel_name	string	Source channel name
platform	string	Source platform
segment_start_sec	float	Start offset in original audio (seconds)
segment_end_sec	float	End offset in original audio (seconds)
speech_duration_seconds	float	Clean speech duration of this segment
sample_rate_hz	int	Native sample rate
channels	int	Audio channels (1=mono, 2=stereo)

Categories

cooking, funny, funny/poetry, funny/stage show, funny/vlog, islamic bayan, nan, news, news/interview , podcast, poetry, telefilms/funny, vlog/daily life/small content/funny, vlog/funny, vlog/memon cooking

Usage

from datasets import load_dataset

ds = load_dataset("hamzahanif/memoni-audio", split="train")
sample = ds[0]
print(sample["unique_id"], sample["speech_duration_seconds"])