Speech recognition
Scripted and spontaneous audio are listed separately, so you can choose the recording conditions and language coverage you need.
Bayanat Voice
Arabic speech datasets for recognition and conversation.
Bayanat Labs (Riyadh) licenses Arabic and Saudi speech datasets for ASR and voice agents: 30,000 hours of Arabic speech, including two separate 1,000-hour Saudi collections and 350 hours of full-duplex customer service.
اقرأ بالعربيةFind your speech dataset
Choose a collection to see its counts, intended use, and licensing details.
Shares a combined 10,000 hours; ask for the split
Arabic Call-Center Conversations
For speech and contact-center teams that need conversation audio rather than individual scripted recordings.
View dataset Shares a combined 10,000 hours; ask for the splitArabic Scripted Monologues
For ASR teams that need scripted recordings with a clear source category for acoustic modeling and recognition experiments.
View dataset 500 hoursArabic Smartphone Speech
For teams training mobile speech recognition or assistants that need to understand Arabic recorded on a phone.
View dataset 1,000 hoursSaudi Spontaneous Speech
For ASR and spoken-language teams that need Saudi speech beyond read scripts.
View dataset 1,000 hoursSaudi Conversational Speech
For teams that need ready-to-license Saudi conversation data with human transcripts and a defined delivery format.
View dataset 350 hoursSaudi Full-Duplex Customer Service
For voice-agent teams working on customer-service interaction and evaluating the labels needed for turn-taking, corrections, and task completion.
View dataset Ask about available volumeSaudi Expressive Speech
For teams developing expressive Arabic voices that need professional recordings and explicit terms for their model use.
View dataset 2,000 prompts · 40 speakers · 2 hoursStudio Prompt Speech
For teams that need a small, clearly defined studio collection to test a speech interaction or explore a voice prototype.
View datasetBayanat Voice totals 30,000 hours of Arabic speech. Call-center conversations and scripted monologues make up a 10,000-hour shared catalog; ask us for each product’s allocation. The two Saudi 1,000-hour collections are separate products. Other collections may overlap, so check the release details before combining them.
Which Saudi or Arabic speech dataset fits your task?
Bayanat lists each collection separately, so you choose by recording type and task rather than a pooled hour total. Saudi rows come first; the last two rows cover UAE and other-variety projects, and synthetic voice.
| Your task | Start with | Check before licensing |
|---|---|---|
| A Saudi customer-service voice agent, or a model handling two-way speech | Saudi Full-Duplex Customer Service (350 hours) | Channels, overlap, interruptions, task labels and transcript coverage, plus the interaction source for the release. Ask for the benchmark report: name, version, evaluated model, metric and score. |
| Saudi conversational ASR or speaker-aware systems that need human transcripts | Saudi Conversational Speech (1,000 hours) | Speaker count, Saudi varieties and collection dates; the schema version and field definitions; a sample with its transcript and metadata. |
| Unscripted Saudi speech for recognition and spoken-language understanding | Saudi Spontaneous Speech (1,000 hours) | Transcript coverage and task labels, speaker count, Saudi varieties and collection dates. This collection is separate from Saudi Conversational Speech. |
| Arabic contact-center speech for ASR and dialogue modeling | Arabic Call-Center Conversations | The hour allocation (10,000 hours are shared with Arabic Scripted Monologues), dialect and speaker mix, call source and recording details. |
| Speech recorded on phones for a mobile assistant | Arabic Smartphone Speech (500 hours, 2025–2026) | Hours in the scripted and spontaneous subsets, speakers and dialects in each, devices and recording conditions. |
| Controlled scripted coverage or acoustic modeling | Arabic Scripted Monologues | The hour allocation and prompt source. Keep scripted audio labeled if you combine it with spontaneous speech. |
| Emirati speech for a UAE project, or another specific variety | Custom speech collection | The Saudi collections are labeled Saudi, not Emirati. Ask for the dialect breakdown of the wider Arabic collections first. If your variety is not covered at the volume you need, scope a collection and name the data location you require. |
| A synthetic voice or expressive text-to-speech | Bayanat Voice Studio Saudi Expressive Speech | Style coverage and speaker permissions. Synthetic voice creation, speaker reuse and distribution need explicit contract scope for every collection, including Saudi Expressive Speech, so name those uses in the contract. |
Is Bayanat Voice the right fit?
A good fit if
- You are training or adapting Arabic speech recognition for phone, contact-center or smartphone audio.
- You need Saudi speech beyond read scripts: spontaneous, conversational or full-duplex customer-service recordings.
- You want to license existing hours now and add annotation, such as speaker segments, overlap or intent, where your model needs it.
Choose something else if
- You need a reusable synthetic voice or studio recordings for TTS. Start with Voice Studio.
- You need dialects or recording conditions these collections do not cover. Commission a custom collection through our speech data service.
What you can check before you license
We separate what is confirmed on this page, what we send on request, and what is agreed for each project.
| Item | Status | Details |
|---|---|---|
| Hours per collection | Confirmed | 30,000 hours of Arabic speech across Bayanat Voice. Listed collections: 500 hours of smartphone speech (2025–2026); 1,000 hours of Saudi spontaneous speech; a separate 1,000 hours of Saudi conversational speech; 350 hours of Saudi full-duplex customer service; 10,000 hours shared by call-center conversations and scripted monologues. |
| Transcripts and delivery schema | Confirmed | Saudi Conversational Speech includes human transcripts, timestamps, speaker attributes and a fixed delivery schema. Other collections: field list on request. |
| Speakers, dialect mix and collection dates | On request | Request them for each collection. |
| Recording source | On request | Natural interactions, role-play, scripted and synthetic audio are kept distinguishable. Ask which applies to each release. |
| Audio format and sample rate | On request | Per collection, with the datasheet. |
| Sample audio | On request | Ask for a sample where one is available, with transcripts where the collection includes them, to check fit before licensing. |
| Benchmark report | On request | Available for Saudi spontaneous and Saudi full-duplex speech: benchmark name, version, evaluated model, metric and score. |
| Overlap between collections | On request | Check the release details before combining collections, and ask about any overlap between them. |
Need more than the existing collection?
Our team can add data or annotation for your particular model. These additions are agreed separately.
Conversation
Compare call-center, Saudi spontaneous, Saudi conversational, and full-duplex collections. Each has its own specifications and hour count.
Additional annotations
Request verbatim and normalized transcripts, speaker segments, timestamps, overlap, language switching, entities, or intent. Included fields vary by collection.
Voice-agent behavior
Add labels for corrections, clarification, confirmations, escalation, and resolution when you need to train beyond transcription.
How the license covers each use
Training rights do not automatically include display, retrieval, redistribution or a synthetic voice. Each use is set in the agreement.
Model training
Commercial license. Permitted training use and delivery scope are set in the agreement. Saudi spontaneous and full-duplex speech include commercial AI-use rights.
Display, retrieval and redistribution
Separate scopes. Name them in your request if your product needs them.
Synthetic voice
Not implied by recognition data rights. It needs explicit speaker terms (see Voice Studio).
How we put the data together
Choose a collection
Review source, dialect, speaker mix, and recording conditions against the conversations your product will handle.
Check the sample
Ask for a sample where one is available, with the transcript and delivery details. Custom deliveries can include original recordings and separately labeled audio derivatives.
Set the terms
Confirm your intended use, required labels, and splits. Real interactions, role-play, and synthetic audio remain distinguishable.
For example: A correction in the middle of a call
A caller changes a building number while the agent is speaking. The useful test is whether the agent hears the correction and saves the right address.
Scope your request
Send us these details and we’ll send the available specifications, the licence wording, and a sample where one is available, so your team can check the fit before licensing. If you need annotation added, we scope a pilot with gold-standard QA and a quality report, usually within two weeks.
- Collection(s) you are interested in
- Task: ASR, diarization or voice agent
- Dialects, channels and recording conditions
- Hours needed
- Uses to license: training, evaluation, display, redistribution
- Timeline
Common questions
Are the two 1,000-hour Saudi speech products the same?
No. Saudi Spontaneous Speech and Saudi Conversational Speech are separate collections, each with 1,000 hours. The conversational product includes human transcripts, timestamps, speaker attributes, and a fixed delivery schema.
Which Saudi Arabic speech dataset should I start with?
It depends on the task. Saudi Full-Duplex Customer Service (350 hours) fits customer-service voice agents and models handling two-way speech, Saudi Conversational Speech (1,000 hours) comes with human transcripts and speaker attributes, and Saudi Spontaneous Speech (1,000 hours) is unscripted. The selection table on this page lists what to check for each.
Do you have Emirati or UAE speech data?
No collection in our catalog is labeled Emirati. Our Saudi collections are labeled Saudi, and the dialect mix of the wider Arabic collections is available on request. If they do not cover the Emirati speech you need, we can scope a custom collection through our speech data service.
Does the 10,000-hour catalog include 10,000 hours of each recording type?
No. That is the combined total for Arabic Call-Center Conversations and Arabic Scripted Monologues. They are separate products; request the hour allocation and speaker mix for each.
Can I use recognition data to create a synthetic voice?
That depends on the collection’s license and speaker permissions. Synthetic voice creation needs explicit terms. Our Studio category covers expressive recordings and custom voice production.