Restricted to NPL1.2 and the Quran Lab team
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is available only to NPL1.2 and the internal Quran Lab team. Requests from anyone else are declined.
Log in or Sign Up to review the conditions and access this dataset content.
Quran Lab Arabic Speech
Broad Arabic speech for next-generation ASR, built from 50,000+ hours of recordings.
Access is restricted to NPL1.2 and the internal Quran Lab team.
Contents
- Broad Arabic speech in segments of 2 to 20 seconds, cut at natural pauses (silences of 1.5 s or longer are removed), sized for CTC/transducer training and for LLM transcription.
- Audio is 48 kHz Opus, exactly as the source streamed it (never re-encoded).
- Full provenance on every segment: source video, channel, title and the segment's position in the original recording.
TTS pool
Config tts (data/tts/) holds studio Arabic podcasts and professional narration. Each recording is graded for TTS
use on four 9 s windows inside its speech, and the grade is in metadata/recordings.parquet:
| column | meaning |
|---|---|
tts_pass |
1 when every check below passes |
tts_reason |
ok, or the first check that failed (narrow_band, music_bed, little_speech, noisy, distorted, low_overall) |
bandwidth_hz |
where the spectrum falls off its cut-off; at least 15 kHz to pass |
dnsmos_sig, dnsmos_bak, dnsmos_ovrl |
DNSMOS P.835; at least 3.4, 3.9 and 3.1 to pass |
Recordings with music under the speech (music score above 0.15) or under 75% speech also fail. Most are conversations with two or more speakers, so they need speaker separation before single-voice training.
Schema
| column | description |
|---|---|
audio |
the segment, Ogg Opus |
segment_id |
<video_id>_<index> |
video_id, segment_index |
source recording and the segment's order in it |
start_s, end_s, duration_s |
position in the original recording |
pool |
content category |
title, channel_id, channel_name, upload_date, url |
source |
video_duration_s |
length of the original recording |
quran_score, lid_ar |
automatic content scores |
Recording metadata
metadata/recordings.parquet has one row per source recording (join on video_id): source, duration, the cleaning
scores, the shard holding its segments, and dup_of. The same recording is often uploaded by several channels;
dup_of names the copy to keep, so rows where it is set can be skipped to avoid training on the same audio twice.
Loading
from datasets import load_dataset
ds = load_dataset("Muno459/QuranLab-Arabic-Speech", split="train", streaming=True)
tts = load_dataset("Muno459/QuranLab-Arabic-Speech", "tts", split="train", streaming=True) # then keep tts_pass == 1
- Downloads last month
- 2,967