500 hours of Arabic smartphone speech
Bayanat Labs (Riyadh) licenses 500 hours of Arabic smartphone speech collected in 2025–2026, with scripted and spontaneous subsets labeled separately, part of Bayanat Voice’s 30,000 hours of Arabic speech.
اقرأ هذه العائلة بالعربيةWhat this dataset is for
For teams training mobile speech recognition or assistants that need to understand Arabic recorded on a phone.
Commercial license; permitted training use and delivery scope agreed in the license.
What’s included
- 500 hours of Arabic smartphone speech
- Collected in 2025–2026
- Separate labels for scripted and spontaneous subsets
- Alignment, evaluation, and agent-trajectory sets available to license with the audio
What Bayanat Labs adds
- One supplier: license this collection from Bayanat Labs (Riyadh), then add custom collection, annotation, RLHF or evaluation from the same team, across nine product families.
- A team of 50,000+ expert annotators, the largest in the Middle East, including vetted native speakers, linguists and licensed domain professionals across 25+ Arabic varieties, calibrated against gold standards with adjudication and an audit trail.
- In-region hosting by default, with on-prem and private-cloud options. We build no models, so we never compete with yours.
- Methods in the open: LabelBench, the Arabic Agent Reliability Lab (methods beta) and our free datasets and tools.
Technical specification: source, format, coverage, limitations & license
Bayanat Labs licenses 500 hours of Arabic smartphone speech, collected in 2025–2026, with scripted and spontaneous subsets labeled separately.
- Source
- Smartphone recordings; scripted and spontaneous subsets labeled separately.
- Record format / annotation
- Smartphone audio with scripted and spontaneous subsets labeled separately; hours per subset, transcripts and alignment on request.
- Audio format
- Audio format, sample rate and channel layout on request; devices and recording conditions on request
- Speakers
- Count and composition on request
- Collection year
- 2025–2026
- Language / dialect
- Arabic; register and dialect breakdown on request
- Quantity
- 500 hours
- Known limitations
- The hour split between subsets is on request. Alignment, evaluation and agent-trajectory sets can be licensed alongside the audio; they are not assumed to be included in the audio license. Retrieval, display and redistribution are separate scopes. Synthetic-voice use is not implied by recognition data rights; it needs explicit speaker terms.
- License type
- Commercial license; permitted training use and delivery scope agreed in the license.
- Refresh cadence
- Ask about versions and updates
- Training task
- Mobile speech recognition and spoken-assistant training
- Last updated
- 2 November 2026
What to check in your sample
- Speakers, dialects, and hours in each subset
- Devices, recording conditions, transcripts, and alignment
- Companion dataset versions and any links to audio records
We’ll send the available sample, specifications and licence wording for this collection so your team can check the fit before licensing. Need it annotated, extended or evaluated on your model? We scope a fixed-scope pilot with gold-standard QA and a quality report, usually within two weeks.
Common questions
Can we distinguish scripted speech from spontaneous speech?
Yes. The subsets are labeled separately. Request the hour split and sample records to choose the balance that fits your training or evaluation.
What other data can we license with the audio?
Arabic alignment, evaluation, and agent-trajectory sets are available alongside the audio. Ask for their versions, formats, and any record-level links; they are not assumed to be included in the audio license.