Translate speech into Arabic or Kurdish on your own laptop, with nothing sent to the internet

What happens to your information

What it costs

Before you start

• A supported managed laptop with at least 8 GB RAM, encryption, microphone access and installation permission.
• Python 3.11 or 3.12 and enough disk and bandwidth for packages and models.
• A public Arabic test script and recording with known text.
• A fluent reviewer for source speech and target language.

The steps

1. Step 1 / 8

   Run the go or no-go preflight: operating system, processor, Python, memory, disk, encryption, microphone, sync, backups and installation permission. Stop on Intel Mac or an unsupported dependency combination.

   Check it worked

   The device and route are recorded as supported, encrypted, non-synchronised and available for a public test.

2. Step 2 / 8

   Create an isolated folder and Python environment. On Linux, install the current CPU-only PyTorch build before the remaining packages to avoid unnecessary accelerator downloads.

   Copy this exactly

   CopyCopied

   python -m venv .venv
   # Windows PowerShell: .\.venv\Scripts\Activate.ps1
   # macOS or Linux: source .venv/bin/activate
   python -m pip install -U pip setuptools wheel

   Check it worked

   The activated environment reports the expected Python and PyTorch versions, and PyTorch reports CPU operation on a non-GPU laptop.

3. Step 3 / 8

   Install RealtimeSTT with its recommended voice detector, CTranslate2, Transformers, SentencePiece and Protobuf. Use current official installation guidance if a platform dependency is missing.

   Copy this exactly

   CopyCopied

   python -m pip install "RealtimeSTT[recommended]" ctranslate2 transformers sentencepiece protobuf
   python -c "import ctranslate2, transformers, faster_whisper, silero_vad; from RealtimeSTT import AudioToTextRecorder; print('packages ready')"

   Check it worked

   Python imports RealtimeSTT, ctranslate2, transformers, faster_whisper and silero_vad without an error.

4. Step 4 / 8

   Download and convert the NLLB translation model, then download the Whisper small model. Use public setup material only.

   Copy this exactly

   CopyCopied

   ct2-transformers-converter --model facebook/nllb-200-distilled-600M --output_dir models/nllb-600m-int8 --quantization int8 --copy_files tokenizer.json tokenizer_config.json special_tokens_map.json sentencepiece.bpe.model
   python -c "from faster_whisper import WhisperModel; WhisperModel('small', device='cpu', compute_type='int8'); print('Whisper ready')"
   python -c "import os, silero_vad; p=os.path.join(os.path.dirname(silero_vad.__file__),'data','silero_vad_op18_ifless.onnx'); print('VAD on disk:', os.path.exists(p))"

   Check it worked

   The NLLB folder exists, the Whisper model is cached, and the installed silero_vad package contains its ONNX model file.

5. Step 5 / 8

   Create the live script from the maintained CDR workshop module. For Arabic speech to Sorani text use ar, arb_Arab and ckb_Arab. Keep model loading inside the main process as documented by the module.

   Copy this exactly

   CopyCopied

   Use the current script from https://github.com/cdrorgtr/cdr-ai/tree/main/simultane-ceviri and save it as translate_live.py. Then run:
   python -m py_compile translate_live.py

   Check it worked

   The script passes a Python syntax check and shows the intended source and target codes before microphone use.

   Note

   NLLB is non-commercial and its model card excludes certified, legal and medical translation uses.

6. Step 6 / 8

   Run the pipeline online with the public known-answer sample. Check microphone capture, source transcript, target translation and response delay.

   Copy this exactly

   CopyCopied

   python translate_live.py

   Check it worked

   A fluent reviewer records errors in names, numbers, negation, omissions, meaning and unacceptable delay.

7. Step 7 / 8

   Stop the program, disconnect Wi-Fi, cable and hotspot, restart the device if practical, and run the same public sample again.

   Copy this exactly

   CopyCopied

   python translate_live.py

   Check it worked

   The pipeline starts and completes without a download or network request while all visible network routes are disconnected.

8. Step 8 / 8

   Make a go or no-go decision. Require a fluent reviewer during operational use, keep original audio where authorised, and record model, settings, language pair, licence, date and reviewer.

   Check it worked

   The work log contains a recovery route to an interpreter or manual transcription.