TECHNICAL SPECIFICATIONS

How we record, transcribe and deliver.

The standard set-up behind our Arabic dialect datasets. Every element can be adapted to your specification: format, conditions, devices, schema and naming.

المواصفات التقنية

SESSIONS

Natural conversation, in controlled
conditions.

Format Conversational speech between 2 speakers (sometimes 3), plus short read prompts. Sessions of 50–60 minutes of recording within a 90-minute slot.
Topics Free conversation guided by our own topic cards (daily life, work, food, travel, family, sport, technology). Sensitive topics and personal details are avoided.
Speakers Adults, native speakers of the dialect, recruited directly. Region of origin checked at screening. Balanced by gender, age band and city.
Code-switching French and English as people naturally mix them into Maghrebi Arabic, captured as spoken.
Environment About 80% in an acoustically treated quiet room; about 20% in controlled real-world noise (café corner, parked car, office). Every file is tagged with its environment. Ratios follow your specification.
Privacy Private room, pseudonymous IDs, no names, addresses or ID numbers spoken. Any personal detail said by mistake is noted and removed.
CAPTURE

Studio-grade capture, one track per
speaker.

Recording room 9–16 m², away from the street, treated with absorbent panels, rug and heavy curtains. Accepted only after a noise and echo test (background around 35 dBA or below, recorded noise floor quieter than −60 dBFS).
Equipment Multitrack field recorder, 32-bit float; one lavalier microphone per speaker on its own track; a backup recorder running in parallel; closed-back headphones for live monitoring.
Capture format WAV, 48 kHz, 32-bit float (or 24-bit), uncompressed, one mono track per speaker. No compression, limiter or noise reduction at capture.
Session markers 10 seconds of silence, spoken session ID and a sync clap at the start of each take.
File naming TN_S0007_SPK014_20261116_01.wav — country, session, pseudonymous speaker ID, date, take.
TRANSCRIPTION

Verbatim, by native speakers of the
dialect.

Script Dialect written verbatim in Arabic script, as spoken (not converted to Modern Standard Arabic). French and English words in Latin script, as spoken.
Segments Time-aligned segments of about 5–20 seconds, cut at natural pauses, with speaker labels.
Tags Non-speech and events marked, for example [laugh] [noise] [overlap], plus unclear words; code-switched spans marked.
Guidelines Written transcription guidelines owned by our linguistic lead; transcribers trained and tested before production.
Quality control Audio checks on every file (clipping, noise, silence, overlap, duplicates), native review, and a second-pass check of a sample of each batch against agreed accuracy targets.
Delivery

Delivered in your format.

  • Audio: WAV, 16-bit PCM, mono, 16 kHz by default; 44.1 / 48 kHz for clean or TTS data; 8 kHz for telephony.
  • Transcripts: JSON, TSV or TextGrid, time-aligned or orthographic.
  • Metadata: JSON or CSV manifest, one line per file.
  • Transfer: secure SFTP or your bucket (S3, Azure, GCS), in weekly batches with a batch report.
  • Your schema: we can adopt your naming, fields and folder structure.
Example manifest line
{
  "file": "TN_S0007_SPK014_20261116_01_seg0047.wav",
  "locale": "ar-TN",
  "region": "Sfax",
  "speaker_id": "SPK014",
  "gender": "F",
  "age_band": "26-40",
  "environment": "quiet_room",
  "duration_s": 11.8,
  "code_switch": true,
  "qa_status": "passed",
  "consent_scope": "commercial_ai_training",
  "tts_opt_in": false
}

Illustrative example. Real samples are shared under NDA.

Need a different set-up?

Far-field, in-car, smartphone, wearable or telephony audio; scripted commands; specific noise conditions. Send us your specification and we will adapt the protocol.