Speech data for AI · Maghreb first
Arabic dialect speech data, recorded where each dialect is spoken.
بيانات صوتية باللهجات العربية
Consented, native-transcribed speech in Tunisian, Libyan and Algerian Arabic — including the natural French–Arabic code-switching real people use — for training and evaluating speech recognition, translation and voice assistants.
آ donc ال projet متاعنا باش نبداو فيه la semaine prochaine
- EU company · GDPR by default
- Written consent from every speaker
- Adults only, paid participation
- Native-speaker transcription and review
- Rejected files replaced at our cost
Why now
Dialects are where speech models still fail.
Models handle Modern Standard Arabic well. Everyday speech in the Maghreb is a different story: dialect vocabulary, no fixed spelling, and constant switching into French. Closing that gap takes real recordings from real speakers, with rights cleared for commercial training.
~36%
average word error rate for the best systems on multi-dialect Arabic speech.
Source: NADI 2025 shared task (arXiv:2509.02038)
36.9%
word error rate on North African Arabic for the strongest open Arabic model tested.
Source: Cohere Transcribe Arabic release, July 2026
Dialect coverage
Maghrebi dialects are our core strength.
Tunisian
ar-TN
Collecting nowLibyan
ar-LY
Collecting nowAlgerian
ar-DZ
Collecting nowMoroccan
ar-MA
On requestEgyptian
ar-EG
On requestLevantine
ar-LB · SY · JO · PS
On requestGulf
ar-AE · SA · KW · QA
On requestModern Standard Arabic
MSA
On requestFrench–Arabic and English–Arabic code-switching is captured as it is spoken and marked in the transcript.
Services
Four services, one Arabic team.
Speech data collection
Scripted, prompted and conversational speech, recorded in-country against your quotas.
Transcription & annotation
Dialect-native orthographic and time-aligned transcripts, with speaker, noise and code-switching tags.
Arabic language services
Translation and translation verification, script verification, error creation and AI output evaluation.
Image data
Image collection and annotation, including Arabic text in images, on request.
Custom scenarios: in-car, call-centre, retail, smart-glasses and wearable commands; far-field or close-talk; real-world noise.
How we work
From your spec to delivered data.
01
Specify
Dialects, quotas, scripts, specs and acceptance criteria.
02
Recruit
Local coordinators recruit native speakers; regional origin checked at screening.
03
Record
Scripted, prompted or conversational, on smartphone, studio or wearable.
04
Review
Native review of every file, then dialect-native transcription.
05
Deliver
Weekly batches by secure SFTP or your cloud bucket.
Quality control. Rejected files are replaced at our cost.
We agree acceptance criteria with you — such as a WER target and a maximum rejection rate — and report against them for every batch.
- Audio checksClipping, background noise, silence, overlap, duplicates.
- Native reviewPrompt adherence and dialect confirmation by a native reviewer.
- Second-pass transcriptionAn independent reviewer checks a sample of each batch.
- Batch reportingHours by dialect, city, gender, age band and rejection rate.
Consent & compliance
Every contributor signs a release for commercial AI training.
Written consent
Covers recording, transfer of rights and AI/ML training. Arabic and English versions. Voice-synthesis use is a separate opt-in.
GDPR by default
Pseudonymised speaker IDs. Names and signed forms stored apart from audio and never delivered. Consent records available for audit.
Ethical sourcing
Adults only, recruited directly — never through an anonymous crowd — and paid fairly and on time.
Technical specs
Delivered in your format.
| Audio | WAV, 16-bit PCM, mono · 16 kHz (44.1 / 48 kHz for clean or TTS data · 8 kHz for telephony) |
|---|---|
| Conditions | Real-world (café, street, car) or studio, as specified |
| Transcripts | Dialect-native, verbatim, orthographic or time-aligned; code-switching marked |
| Metadata | JSON or CSV: speaker ID, gender, age band, dialect, region, device, environment, duration |
| Delivery | Secure SFTP or your bucket (S3, Azure, GCS), weekly batches · we can adopt your schema and naming |
Working with us
Start with a pilot. Pay once you have validated it.
STEP 1
Free sample
30 minutes of Tunisian speech with manifest and speaker sheet, to test against your spec.
No charge · under NDA
STEP 2
Pilot
About 10 hours, one dialect, ~20 speakers. Starts within 2 weeks of agreed specs; delivered in 3–4 weeks. No minimum order.
Fixed price, quoted per spec
STEP 3
Production
80–120 hours of audio per month, or 100–200 participants, balanced by gender, age and city.
Per accepted audio hour
CUSTOM
To your spec
Your prompts, domain, devices and noise conditions, on your platform or ours.
Per produced hour
Licensing: exclusive with full IP transfer, time-limited exclusive, or non-exclusive. Volume orders invoiced per accepted batch.
Leadership
Project discipline and native linguistics.
Mourad Ben Chedli
Founder · Project lead
PMP-certified, Master's in Project Management (Université du Québec). 15+ years running projects and operations in Canada, France, Belgium, Tunisia and the UAE, including IT projects for the Government of Canada.
Ichrak Gharbi
Linguistic lead
Arabic language teacher and graduate of the University of Manouba (Arabic language and literature). Owns the transcription guidelines and dialect quality control.
Let's scope your Arabic pilot.
Tell us the dialects, hours and use case. We reply within one business day. Samples and the full specification are shared under NDA.
Email
contact@arabicaivoice.com
Phone
+216 23 22 22 66 (Tunisia)
+971 50 337 1701 (UAE)
+33 6 68 02 80 77 (France)
Offices
Brussels · Tunis · Dubai