Question 1

How accurate is CallScribe for Gulf Arabic?

Accepted Answer

CallScribe uses Whisper large-v3-turbo optimized for Arabic dialects. Internal testing across 200+ Gulf Arabic call recordings (March 2026) shows 85-95% word-level accuracy for Khaleeji dialect with clear audio (SNR > 15dB).

Question 2

Is my call data private with CallScribe?

Accepted Answer

Yes. CallScribe processes everything on your own server. No audio, transcripts, or metadata are sent to external servers. Zero external API calls during transcription.

Question 3

What file formats does CallScribe support?

Accepted Answer

CallScribe accepts MP3, WAV, M4A, FLAC, OGG, and WebM audio files. Export formats include PDF, CSV, TXT, DOCX, and SRT for subtitle workflows.

Question 4

How much does CallScribe cost?

Accepted Answer

CallScribe offers a free Starter plan (5 min/month), Business plan at $29/month (500 min), and Scale plan at $79/month (3000 min). No per-minute charges.

Question 5

Does CallScribe support Khaleeji dialect?

Accepted Answer

Yes. CallScribe uses Whisper large-v3-turbo optimized for Gulf Arabic (Khaleeji). Internal testing across 200+ call recordings shows 85-95% word-level accuracy with clear audio (SNR > 15dB). March 2026 benchmark.

Question 6

Can CallScribe handle code-switching between Arabic and English?

Accepted Answer

Yes. CallScribe detects when speakers switch between Arabic and English mid-sentence — a common pattern in GCC business calls. Both languages are transcribed accurately.

Question 7

Is CallScribe GDPR compliant?

Accepted Answer

Yes. All processing happens on private infrastructure hosted in the EU (Hetzner, Germany) with optional GCC residency via Tailscale-routed workers. No audio, transcripts, or metadata are sent to external US-based servers during transcription. CallScribe publishes a Data Processing Agreement (DPA) covering GDPR Article 28 processor obligations, sub-processor disclosures (Stripe for billing, Resend for transactional email, Sentry for error telemetry), 72-hour breach notification, and data deletion on termination. The platform is also aligned with UAE PDPL data sovereignty requirements. Data Subject Access Requests can be submitted to privacy@callscribe.ae and are answered within 30 days.

Question 8

What dialects of Arabic does CallScribe transcribe most accurately?

Accepted Answer

CallScribe is tuned primarily for Khaleeji (Gulf) Arabic — including Emirati, Saudi, Kuwaiti, Qatari, Bahraini, and Omani variants — where internal benchmarks on 200+ clear-audio calls (SNR > 15 dB) show 85-95% word-level accuracy. Levantine Arabic (Lebanese, Syrian, Palestinian, Jordanian) lands around 84-88%. Egyptian Arabic reaches 86-90%. Modern Standard Arabic (MSA), common in news and formal recordings, reaches 91-94%. Maghrebi dialects (Moroccan, Algerian, Tunisian) are not officially supported in this release. Accuracy drops on heavily overlapping speakers, poor SNR below 10 dB, or long-form mumbling. See the /model-card page for the full methodology and per-dialect WER table.

Question 9

How does CallScribe compare to AWS Transcribe and Google Speech-to-Text for Arabic?

Accepted Answer

AWS Transcribe and Google Speech-to-Text both support Arabic but are tuned primarily toward Modern Standard Arabic, with limited coverage of Gulf, Levantine, and Egyptian dialects. In internal comparisons on Khaleeji call center recordings, both vendors produced 15-25% higher word error rates than CallScribe's Whisper large-v3-turbo pipeline. CallScribe also processes audio on private infrastructure in the EU or on GCC-resident workers — AWS and Google route audio through US regions by default, which is a compliance blocker for UAE PDPL and many GCC enterprise procurement policies. Pricing is flat-rate per minute bucket instead of per-second billing, which tends to save 30-60% at call center volume.

Question 10

Can I deploy CallScribe on my own infrastructure?

Accepted Answer

Yes. CallScribe offers a self-hosted deployment option for Scale-tier customers and enterprise buyers. The stack runs entirely on Docker Compose: a Fastify API, a Python worker running Whisper large-v3-turbo and pyannote.audio for diarization, PostgreSQL, Redis, and nginx. Typical hardware: a single GPU worker (RTX 4090 or L4) handles up to 10x realtime throughput. The API tier runs comfortably on a 4-core VPS. Tailscale is used to connect worker nodes back to the control plane over a private mesh, so the GPU host can live in your own rack while the API stays in the cloud. Contact sales@callscribe.ae for a self-host deployment guide and license terms.

Question 11

What is the turnaround time for a 1-hour call?

Accepted Answer

On the shared Business tier, a 1-hour call typically completes in 4-8 minutes end-to-end — including upload, transcription with Whisper large-v3-turbo, speaker diarization via pyannote, sentiment analysis, and audio quality scoring. Scale tier customers get priority queue placement and usually see 2-4 minutes for the same file. Self-hosted deployments on an RTX 4090 consistently process 60 minutes of audio in under 3 minutes (over 20x realtime). Queue wait time is the largest variable during peak hours on the free tier. WebSocket progress updates stream live from the worker so users see per-file percent-complete rather than a silent spinner.

Question 12

How does CallScribe handle noisy call center recordings?

Accepted Answer

Call center audio is rarely clean — hold music bleed, codec artifacts, echo, cross-talk, and background PA announcements are all common. CallScribe runs a pre-processing pipeline that analyzes signal-to-noise ratio (SNR), root-mean-square loudness, and speech activity before transcription, then reports a per-file audio quality score so users know how much to trust the transcript. For SNR above 15 dB accuracy stays in the 85-95% band. Between 10 and 15 dB it degrades to the high 70s. Below 10 dB, CallScribe flags the file as low-confidence and recommends re-recording or applying an external denoiser before retry. Overlapping speakers are split via pyannote diarization, not just channel separation, so mono call recordings still work.

Create account