Arabic Transcription With Whisper on Your Server: Full Privacy, Fixed Cost
Running Whisper on a Saudi server transcribes your meetings, minutes, and support calls in Arabic without the audio ever leaving your infrastructure. Pick a model size by RAM and speed, and weigh a fixed annual cost against per-minute external APIs.
If you want to transcribe Arabic audio (meetings, minutes, support calls) without the recording ever leaving your server, run the open-source Whisper model on a Saudi cloud server through the faster-whisper library. The result: recordings stay inside your infrastructure in Riyadh, cost becomes a fixed annual figure instead of a per-minute charge to an external API, and accuracy on clean Modern Standard Arabic reaches a word error rate under 15% with the large-v3 model.
The practical decision in one line: for personal use or a small team, the small model on a Growth cloud server with 4GB RAM is enough. For high accuracy on Arabic, use large-v3 on an Enterprise plan with 16GB RAM or add memory. A call center transcribing thousands of hours a month saves real money by self-hosting, because cost stops at the annual server rent no matter how much audio grows.
This guide explains Whisper's model sizes and their memory and speed needs, why faster-whisper is much quicker than the reference implementation, then works out the cost per hour of audio on a VPS versus an external API, and closes with real Saudi use cases.
Whisper model sizes: memory versus accuracy and speed
Whisper ships in five core sizes, from tiny at 39 million parameters to large at 1.55 billion parameters. Each jump up improves accuracy, raises memory use, and slows transcription. On Arabic specifically, the gap between small and large is wide, because the language leans on diacritics and context that larger models capture better.
The table below summarizes requirements when running on the CPU in int8 through faster-whisper, the practical format on a cloud server with no GPU:
| Model | Parameters | Approx. memory (int8) | Plan fit | Arabic accuracy |
|---|---|---|---|---|
| tiny | 39M | ~1GB | Basic 2GB | Weak, for testing |
| base | 74M | ~1GB | Basic 2GB | Acceptable for clear speech |
| small | 244M | ~2GB | Growth 4GB | Good for MSA |
| medium | 769M | ~5GB | Business 8GB | Very good |
| large-v3 | 1550M | ~8 to 10GB | Enterprise 16GB | Best, under 15% error on MSA |
An important note on dialects: the numbers above are for clean Modern Standard Arabic. Recordings in local dialects or with background noise push the error rate to between 25% and 40% even with large-v3. That is why we always suggest testing a sample of your own real audio before committing to a size, since microphone and recording quality sometimes matter more than model size.
Why faster-whisper and not the reference implementation
The faster-whisper library rebuilds Whisper on top of the CTranslate2 engine, giving roughly four times the speed and lower memory use at the same accuracy. The gain comes from techniques such as int8 quantization, which shrinks the model in memory by half or more with no noticeable loss in Arabic quality.
On a modern eight-core CPU (Enterprise plan), the medium model transcribes an hour of audio in about real time or less, while large-v3 needs longer but stays practical for overnight batch processing. If you need live transcription during a call, small or medium is the better choice, because large-v3 on CPU alone may not keep pace with live speech.
The short setup steps
- Install Python and the faster-whisper package on the cloud server, or use a ready Docker template from AI hosting.
- Download the model once (it is cached locally), then pass the audio file for transcription with the Arabic language set explicitly to improve accuracy.
- Enable voice activity detection (VAD) to skip silence, which cuts processing time on long calls with many silent gaps.
- Save the output as timed text (SRT or VTT) if you need subtitles or time-anchored minutes.
Cost math: one hour of audio on a VPS versus an external API
The common external API prices transcription at around 0.006 dollars per minute, which is 0.36 dollars per hour of audio, roughly 1.35 SAR per hour. It looks cheap at first, but the cost grows linearly with audio volume, and the recording leaves your infrastructure to a server outside the Kingdom, which alone is a blocker for many sensitive entities.
Self-hosting, by contrast, has a fixed cost: you pay the annual server rent no matter how much audio grows. The table compares the two across usage levels:
| Monthly usage | External API (1.35 SAR/hour) | Business server (11,500 SAR/year) |
|---|---|---|
| 200 hours | ~3,240 SAR/year | 11,500 SAR fixed |
| 1,000 hours | ~16,200 SAR/year | 11,500 SAR fixed |
| 3,000 hours | ~48,600 SAR/year | 11,500 SAR fixed |
The reading is clear: at small volumes the external API looks cheaper, but the break-even point sits near 700 hours a year, after which self-hosting is far cheaper. On top of that, the server does not serve transcription alone, it also runs databases and other apps alongside, so the cost is shared. For plans and add-ons see pricing and plans, since a RAM upgrade of 600 SAR per 2GB is usually enough to move from medium to large.
Raising Arabic accuracy in practice
Model size is not the only factor in accuracy. Several technical steps lift output quality on Arabic noticeably without changing the plan:
- Set the language explicitly: pass the Arabic language code ar rather than letting the model auto-detect, since auto-detection can confuse Arabic with Persian or Urdu on short clips and ruin the start of the transcript.
- Give a priming context: Whisper accepts an initial prompt that steers it toward expected terms, so if the meeting is technical, feed it examples of the recurring names and terms so it writes them correctly instead of guessing.
- Split long audio: calls beyond two hours are better cut into shorter segments using voice activity detection, which reduces error accumulation and uses cores in parallel.
- Clean the audio first: noise reduction and level normalization can raise accuracy more than a full jump in model size, especially for field recordings.
If you need very fast live transcription, the CPU alone may not keep up with large-v3, and this is where a GPU accelerates transcription tenfold or more. Most business cases do not need that, since overnight batch processing solves the speed question without extra hardware cost, but it is worth knowing the limit before choosing a plan.
Real Saudi use cases
Meeting minutes
A government entity or company records its meetings and wants a text record without uploading audio to any external service. Self-hosted transcription produces a full timed transcript, which can then pass to a local language model for summarizing and extracting decisions, all inside Riyadh at a response time near 8ms.
Support center calls
A call center transcribes thousands of calls to analyze service quality and spot recurring complaints. Here self-hosting is practically mandatory, because calls carry sensitive customer data, and the large volume makes per-minute billing very expensive.
Content and media
A content producer transcribes podcast episodes or clips to produce subtitles and searchable, archivable text. The medium model is usually enough, and overnight batch processing uses the server without interrupting daytime work.
Data sovereignty is not a detail
Transcribed audio usually carries people's names, contract numbers, and decision details, and this data should not pass through a server outside the Kingdom just to save an install step. When the recording and its transcription stay inside a Saudi server in Riyadh, the data stays under your full control, you decide who reaches it, you log every operation, and you delete it whenever you want without relying on a third party's policy. This difference alone is enough to tip the balance toward self-hosting for government entities and companies that treat their customers' data seriously, regardless of the cost math.
Conclusion: where to start
Start small: set up a Growth plan, install faster-whisper with the small model, run a sample of your real audio, and measure the error rate on your own speech rather than on generic figures. If you need higher accuracy, move to medium then large-v3, and add RAM when needed at 600 SAR per 2GB. If your monthly volume exceeds 700 hours, self-hosting is cheaper in money terms as well as being the only acceptable option for sensitive data. Try the ready AI hosting template to shorten the install steps, then scale up once you trust the quality.
