ترجم الكلام المنطوق إلى العربية أو الكردية على حاسوبك، دون إرسال أي شيء إلى الإنترنت

ماذا يحدث لمعلوماتك

التكلفة

قبل أن تبدأ

• A supported laptop. Windows 10/11, or an Apple Silicon Mac (M1 or later) running macOS 14 or later, or Ubuntu 22.04+. INTEL MACs CANNOT RUN THIS AT ALL - see the platform notes before you start, because there is no workaround and you will otherwise hit a wall at step 14.
• A laptop with at least 8 GB of RAM. 4 GB will not work. Two of the fourteen organisations surveyed are blocked by hardware and this recipe is one they will be blocked on.
• Free disk space during setup: at least 12 GB on Windows and macOS. On Ubuntu, 12 GB if you do step 13, or 18 GB if you skip it. You get roughly 6 GB back at step 18. Expect to end up using about 4 GB permanently.
• A working microphone. A cheap USB headset microphone will improve accuracy more than any software setting in this recipe.
• Full-disk encryption, turned on before you put anything sensitive through this. Step 2 does it. On Windows Home editions BitLocker may be unavailable, in which case use Device Encryption if your machine offers it - and if it offers neither, do not use this laptop for material that would endanger anyone if the machine were seized.
• Internet once, for the setup only, to download about 4 GB. It does not need to be fast, but it needs to survive a large download - see the USB-stick workaround in the cost note if it will not.
• Permission to install software on the laptop. On a locked-down or managed machine you may not be able to install Python.
• About 100 minutes, most of it waiting for downloads. You can walk away during steps 14, 16 and 17.

الخطوات

1. خطوة 1 / 23كل الأنظمة

   Check your laptop has enough memory, and check which kind of Mac you have if you are on a Mac. ON WINDOWS: press the Windows key, type 'About your PC', press Enter, and read 'Installed RAM'. ON macOS: click the Apple menu, then About This Mac, and read both the Memory line and the Chip line. ON UBUNTU: open Settings, then About.

   تأكد من أنها نجحت

   You need 8 GB or more of RAM. On a Mac you additionally need the Chip line to say Apple M1, M2, M3 or M4 (NOT 'Intel'), and the macOS version to be 14 or higher. If all of that is true, continue.

   ملاحظة

   STOP HERE, AND DO NOT SPEND THE AFTERNOON, IF: it says 4 GB of RAM (no setting will fix this); or your Mac's chip line says Intel (PyTorch has published no Intel-Mac build since March 2024, so the install cannot complete); or your Apple Silicon Mac is on macOS 12 or 13 (update to 14+ first). In any of those cases use the 'someone else runs it and shares the output' route instead, or borrow a Windows laptop.

2. خطوة 2 / 23كل الأنظمة

   Turn on full-disk encryption, before you put anything sensitive through this tool. This is the step that makes 'it stays on my laptop' actually mean something, because laptops get seized, stolen and repaired. ON WINDOWS: press the Windows key, type 'Manage BitLocker', press Enter, and switch BitLocker on for your C: drive; save the recovery key somewhere that is not the same laptop. ON macOS: Apple menu, System Settings, Privacy & Security, scroll to FileVault, and turn it on. ON UBUNTU: whole-disk LUKS encryption can only be chosen while installing Ubuntu; if you did not choose it then, you cannot add it now without reinstalling, so treat this laptop as unencrypted and plan accordingly.

   تأكد من أنها نجحت

   Windows: the BitLocker panel says 'BitLocker on' for your C: drive. macOS: the FileVault panel says 'FileVault is turned on'. Encryption may then run in the background for an hour or more; you can carry on with the recipe while it does.

   ملاحظة

   Why this is in a translation recipe: an 8 GB laptop pages memory out to a swap file, and fragments of the audio buffer and of the decoded transcript can land there and survive after you close the program. macOS encrypts swap by default; many Ubuntu installs do not; on Windows the page file is covered once BitLocker is on. This one step closes the most likely place sensitive content actually ends up on disk. If BitLocker is missing on Windows Home, look for 'Device encryption' in Settings, Privacy & security.

3. خطوة 3 / 23كل الأنظمة

   Open a terminal window. This is the black or white text window you will type every command into. ON WINDOWS: press the Windows key, type PowerShell, and press Enter. ON macOS: hold Command and press the space bar, type the word Terminal, and press Enter. ON UBUNTU: hold Ctrl and Alt and press T.

   تأكد من أنها نجحت

   A window opens with a blinking cursor waiting for you to type. That is all you need.

   ملاحظة

   Nothing you type here is sent anywhere. It is just a way of giving your own computer instructions. When a command below is on more than one line, copy the whole block at once.

4. خطوة 4 / 23كل الأنظمة

   Find out whether Python is already installed and which version. Copy the line for your system, paste it into the terminal, and press Enter.

   انسخ هذا كما هو

   WINDOWS:
   python --version
   
   macOS / UBUNTU:
   python3 --version

   تأكد من أنها نجحت

   You need it to print a version starting with 3.11 or 3.12 - for example 'Python 3.12.7'. If it does, skip step 5 entirely and go to step 6.

   ملاحظة

   If it says 3.13, 3.14 or higher, that is TOO NEW: RealtimeSTT 1.1.2 declares its supported Python range as 3.11 up to but not including 3.13, and pip will simply refuse to install it. If it says 3.10 or lower, or 'command not found' or 'not recognized', it is too old or missing. In all of those cases do step 5. Having a newer Python already installed is not a problem - step 5 installs 3.12 alongside it without removing anything.

5. خطوة 5 / 23كل الأنظمة

   ONLY IF step 4 did not print 3.11 or 3.12: install Python 3.12. Open https://www.python.org/downloads/ in your browser, scroll down to 'Looking for a specific release?', find the most recent release whose number starts with 3.12, click it, scroll to the bottom, and download the installer for your system (Windows installer 64-bit, or macOS 64-bit universal2 installer). Run it.

   انسخ هذا كما هو

   https://www.python.org/downloads/

   تأكد من أنها نجحت

   Close the terminal window completely, open a new one, and run step 4 again. It must now print a version starting with 3.12. If it still does not, the PATH box below is almost always the reason on Windows.

   ملاحظة

   ON WINDOWS THIS MATTERS MORE THAN ANYTHING ELSE ON THIS PAGE: on the very first installer screen, tick the box that says 'Add python.exe to PATH' before you click Install Now. It is easy to miss and if you miss it, none of the later commands will work and the errors will not tell you why. If you already clicked through without ticking it, run the installer again, choose Modify, and complete it - or simply uninstall and reinstall with the box ticked.

6. خطوة 6 / 23ماك

   macOS ONLY: install Homebrew, the tool macOS needs in order to get microphone support. Skip this step entirely on Windows and Linux. Paste the line below and press Enter, then follow the prompts - it will ask for your Mac login password, and the password will not appear on screen as you type it, which is normal.

   انسخ هذا كما هو

   /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

   تأكد من أنها نجحت

   When it finishes, close the terminal, open a new one, and run: brew --version It should print something like 'Homebrew 4.x.x'. If it says 'command not found', you missed the two 'Next steps' lines described below.

   ملاحظة

   UNDERSTAND WHAT YOU ARE AUTHORISING HERE, because this is a pattern criminals also use. This command downloads a script from Homebrew's official address and runs it immediately, with your admin password. It is the genuinely standard way to install Homebrew and it is safe from this exact address - but the shape of it ('paste this line, give it your password') is the shape of many attacks. The rule to carry away: only ever do this for software you went looking for, from the vendor's own documented address that you typed or verified yourself. Never because a message, an email or a stranger told you to. This takes 5-15 minutes. At the end it may print two 'Next steps' lines beginning with 'echo' and 'eval' - copy and run those two lines exactly as printed, then close and reopen the terminal.

7. خطوة 7 / 23ماك

   macOS ONLY: install the microphone library. Skip on Windows and Linux.

   انسخ هذا كما هو

   brew install portaudio

   تأكد من أنها نجحت

   It ends without a red error, and 'brew list portaudio' prints a list of files. If it says portaudio is already installed, that is fine.

   ملاحظة

   Without this, the pip install in step 14 fails with an error mentioning 'portaudio.h'. This is the single most common macOS failure in this recipe.

8. خطوة 8 / 23لينكس

   UBUNTU/LINUX ONLY: install the system packages for microphone and audio. Skip on macOS and Windows. It will ask for your login password.

   انسخ هذا كما هو

   sudo apt-get update && sudo apt-get install -y python3-dev python3-venv portaudio19-dev ffmpeg libsndfile1

   تأكد من أنها نجحت

   The last line of output should be something like 'Processing triggers...' with no 'E:' error lines. If it reports unmet dependencies, run 'sudo apt-get update' on its own first and try again.

   ملاحظة

   Windows and macOS need no equivalent of this step. Windows users go straight to step 9.

9. خطوة 9 / 23كل الأنظمة

   Make a folder to keep everything in, and move into it. Copy the two lines for your system and paste them together.

   انسخ هذا كما هو

   WINDOWS PowerShell:
   mkdir $HOME\live-translate
   cd $HOME\live-translate
   
   macOS / UBUNTU:
   mkdir -p ~/live-translate
   cd ~/live-translate

   تأكد من أنها نجحت

   Your terminal prompt should now show that you are inside the live-translate folder. To be sure, run 'pwd' on macOS/Ubuntu or 'pwd' on Windows PowerShell - it must end in live-translate.

   ملاحظة

   On Windows the folder appears at C:\Users\YourName\live-translate. From here on, every command assumes you are inside this folder. If you close the terminal and come back later, move back into it first with 'cd $HOME\live-translate' on Windows or 'cd ~/live-translate' on macOS and Ubuntu. If mkdir says the folder already exists, that is fine - just run the 'cd' line.

10. خطوة 10 / 23كل الأنظمة

    Create a private workspace for the software so it cannot disturb anything else on your laptop.

    انسخ هذا كما هو

    WINDOWS:
    python -m venv .venv
    
    macOS / UBUNTU:
    python3 -m venv .venv

    تأكد من أنها نجحت

    It takes a few seconds and prints nothing. Confirm it worked by listing the folder: run 'dir' on Windows or 'ls -a' on macOS/Ubuntu, and check that a '.venv' folder is now there.

    ملاحظة

    If you ever want to undo this whole recipe, deleting the live-translate folder removes everything except the Python installation itself.

11. خطوة 11 / 23كل الأنظمة

    Switch the terminal into that workspace. Use the line for your system.

    انسخ هذا كما هو

    WINDOWS PowerShell:
    .\.venv\Scripts\Activate.ps1
    
    macOS / UBUNTU:
    source .venv/bin/activate

    تأكد من أنها نجحت

    '(.venv)' must appear at the start of your terminal line. If it does not, the next steps will install software into the wrong place. Do not continue until you see it.

    ملاحظة

    IF WINDOWS REFUSES with a message about 'running scripts is disabled on this system', run this first and then try the activation line again: Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass What that does, in plain terms: Windows blocks script files by default as a safety measure, and this lifts the block for THIS TERMINAL WINDOW ONLY - the '-Scope Process' part means it forgets the moment you close the window, and it changes nothing permanently on your computer. It is the normal way to activate a Python workspace. Be wary of any instruction that tells you to run the same command WITHOUT '-Scope Process', or with '-Scope LocalMachine', because that would lower the setting for your whole computer permanently. YOU MUST DO THIS ACTIVATION STEP EVERY SINGLE TIME you open a new terminal to use this tool.

12. خطوة 12 / 23كل الأنظمة

    Update the installer tool itself, so the next steps do not fail on an old version.

    انسخ هذا كما هو

    python -m pip install -U pip setuptools wheel

    تأكد من أنها نجحت

    It ends with a line beginning 'Successfully installed' (or 'Requirement already satisfied' for all three). Takes under a minute.

    ملاحظة

    Note it is now 'python', not 'python3' - inside the activated workspace, 'python' is the right one on all systems. If this errors with 'no such file or directory', you are not inside the workspace: go back to step 11.

13. خطوة 13 / 23لينكس

    UBUNTU/LINUX ONLY, AND DO NOT SKIP IT: install the CPU-only version of PyTorch first. Skip this step on Windows and macOS. This single line saves you roughly 4 GB of download.

    انسخ هذا كما هو

    python -m pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu

    تأكد من أنها نجحت

    Run: python -c "import torch; print(torch.__version__)" It must print a version ending in '+cpu', for example '2.13.0+cpu'. If it prints a version with no '+cpu' on the end, the plain version was installed instead - uninstall it with 'python -m pip uninstall -y torch torchaudio' and run this step again.

    ملاحظة

    WHY THIS EXISTS: on Linux, and only on Linux, the ordinary PyTorch package unconditionally drags in several NVIDIA CUDA graphics-card libraries - I checked the current release and nvidia-cudnn-cu13, nvidia-nccl-cu13, nvidia-cusparselt-cu13, nvidia-nvshmem-cu13 and triton are all marked as required whenever the system is Linux. Your laptop almost certainly has no NVIDIA graphics card, and this recipe runs entirely on the CPU anyway, so those gigabytes are pure waste on a metered connection. The Linux CPU-only build is 192 MB against 527 MB plus the CUDA packages. Doing this BEFORE step 14 means step 14 finds PyTorch already present and does not fetch its own copy.

14. خطوة 14 / 23كل الأنظمة

    Install the software: the microphone-and-listening tool with its voice detection, the translation engine, and the helper libraries. This is the longest download of the recipe. Leave it running.

    انسخ هذا كما هو

    python -m pip install "RealtimeSTT[recommended]" ctranslate2 transformers sentencepiece protobuf

    تأكد من أنها نجحت

    It must end with a long line beginning 'Successfully installed' that includes realtimestt, faster-whisper, silero-vad, onnxruntime and ctranslate2. Step 15 checks the critical part properly - do not skip it.

    ملاحظة

    Expect 10-40 minutes depending on your connection. THE WORD 'recommended' IS LOAD-BEARING AND MUST NOT BE CHANGED. An earlier version of this recipe said RealtimeSTT[faster-whisper] instead. That option installs the speech recogniser but NOT the voice-detection model, and RealtimeSTT then silently downloads that model from github.com every time you start the program - which would have made this recipe's central promise false. 'recommended' installs faster-whisper 1.2.1 plus silero-vad with its CPU ONNX runtime, and the silero-vad package carries its own model file inside it, so nothing is fetched later. If the install stops with a red error mentioning 'portaudio.h', you skipped step 7 (macOS) or step 8 (Ubuntu). If it stops with a network timeout, run the identical command again - it resumes rather than starting over.

15. خطوة 15 / 23كل الأنظمة

    Prove the voice-detection model is really on your disk. This is the check that makes the offline promise true rather than hopeful. Paste the whole line.

    انسخ هذا كما هو

    python -c "import os, silero_vad, onnxruntime; p=os.path.join(os.path.dirname(silero_vad.__file__),'data','silero_vad_op18_ifless.onnx'); print('VAD model on disk:', os.path.exists(p)); print('onnxruntime version:', onnxruntime.__version__)"

    تأكد من أنها نجحت

    It must print exactly 'VAD model on disk: True' and then an onnxruntime version number. If it prints 'False', or fails with 'No module named silero_vad' or 'No module named onnxruntime', then step 14 did not install the right thing - re-run step 14 and check you typed [recommended] and not [faster-whisper].

    ملاحظة

    What you just checked: RealtimeSTT looks for this exact file inside the installed silero_vad package before it falls back to downloading a voice-detection model from the internet. Because the file is there, the fallback never runs. If this prints False and you continue anyway, your program will quietly contact github.com every time it starts, and no offline setting in the script will stop it.

16. خطوة 16 / 23كل الأنظمة

    Download the translation model and shrink it so it fits comfortably in memory. This downloads about 2.5 GB and writes out a much smaller compressed model. Paste the whole block as one command.

    انسخ هذا كما هو

    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

    تأكد من أنها نجحت

    When it finishes, list the output folder - 'dir models\nllb-600m-int8' on Windows, 'ls -lh models/nllb-600m-int8' on macOS/Ubuntu. You must see a 'model.bin' of roughly 600 MB (about 623 MB), plus tokenizer.json, tokenizer_config.json, special_tokens_map.json and sentencepiece.bpe.model. IF model.bin IS MISSING OR IS ONLY A FEW KILOBYTES, the conversion failed even if the command looked calm - delete the whole models folder and run this step again. A folder that exists but is nearly empty is the failure to watch for here.

    ملاحظة

    Expect 10-30 minutes. This is NLLB-200 distilled 600M, the model that gives you Kurdish Sorani (ckb_Arab), Kurmanji (kmr_Latn) and Iraqi Arabic (acm_Arab). The '--quantization int8' part is what makes it run on an 8 GB laptop: it turns a 2.46 GB model into roughly 623 MB. Converting it yourself rather than downloading someone's pre-converted copy means you know exactly what you are running. Do NOT delete any download cache yet - step 18 tells you exactly what is safe to delete and when.

17. خطوة 17 / 23كل الأنظمة

    Download the listening model. Paste the whole line.

    انسخ هذا كما هو

    python -c "from faster_whisper import WhisperModel; WhisperModel('small', device='cpu', compute_type='int8'); print('listening model ready')"

    تأكد من أنها نجحت

    It must print 'listening model ready' as the last line. That fetches about 484 MB. If it ends with an error mentioning a connection or a repository, your download did not finish - run the identical line again.

    ملاحظة

    'small' is the right starting size for an 8 GB laptop. To try the more accurate but heavier 'medium' (1.53 GB) later, run the same line with 'medium' in place of 'small' - read the model tier table in step 22's note first.

18. خطوة 18 / 23كل الأنظمة

    OPTIONAL, and only worth doing if you are short of disk space: reclaim about 2.5 GB by deleting the NLLB download cache. You already converted that model in step 16, so this copy is no longer needed. Use the line for your system EXACTLY as written.

    انسخ هذا كما هو

    WINDOWS PowerShell:
    Remove-Item -Recurse -Force "$HOME\.cache\huggingface\hub\models--facebook--nllb-200-distilled-600M"
    
    macOS / UBUNTU:
    rm -rf ~/.cache/huggingface/hub/models--facebook--nllb-200-distilled-600M

    تأكد من أنها نجحت

    Immediately afterwards, re-run the step 17 line. It must still print 'listening model ready' WITHOUT downloading anything again. If it starts downloading 484 MB again, you deleted too much - let it finish and do not repeat this step.

    ملاحظة

    NOTE THE PRECISE PATH: it ends with the named nllb folder. An earlier version of this recipe told you to delete the whole '~/.cache/huggingface/hub' directory, which is also where your Whisper listening model lives - doing that after step 17 silently destroys the listening model and produces an offline failure that is very hard to diagnose. This version deletes only the one folder that is genuinely finished with, so it is safe whenever you run it. If you are not short of space, just skip this step; keeping the cache costs you nothing but disk.

19. خطوة 19 / 23كل الأنظمة

    Create the program file and open it for editing. Use the command for your system.

    انسخ هذا كما هو

    WINDOWS:
    notepad translate_live.py
    
    macOS / UBUNTU:
    nano translate_live.py

    تأكد من أنها نجحت

    A text editor opens. On Windows, Notepad first asks whether to create the file - click Yes. You should be looking at an empty document.

    ملاحظة

    You now have an empty text editor open, waiting for the program in the next step.

20. خطوة 20 / 23كل الأنظمة

    Copy the entire block below - all of it, from the first line to the last - and paste it into the editor you just opened. Then save: in Notepad, hold Ctrl and press S, then close the window. In nano, hold Ctrl and press O, press Enter, then hold Ctrl and press X.

    انسخ هذا كما هو

    # Live speech translation that never leaves your laptop.
    # Adapted from CDR's 'simultane-ceviri' workshop module (November 2024, MIT licence),
    # rebuilt for 2026 with NLLB-200 in place of M2M100 so that Kurdish Sorani works at all.
    
    import os
    
    # Refuse to touch the Hugging Face network. The voice-detection model is loaded
    # from the installed silero-vad package, which ships its own model file, so
    # nothing reaches the network at run time.
    os.environ["HF_HUB_OFFLINE"] = "1"
    os.environ["TRANSFORMERS_OFFLINE"] = "1"
    
    import ctranslate2
    import transformers
    from RealtimeSTT import AudioToTextRecorder
    
    # ---------------- CHANGE THESE THREE LINES TO PICK YOUR LANGUAGES ----------------
    LISTEN_LANGUAGE = "ar"        # the language you SPEAK, in Whisper's code
    NLLB_SOURCE     = "arb_Arab"  # the SAME language, in NLLB's code
    NLLB_TARGET     = "ckb_Arab"  # the language you want to READ on screen
    # ---------------------------------------------------------------------------------
    
    WHISPER_MODEL = "small"                  # tiny | base | small | medium | large-v3
    MODEL_DIR     = "models/nllb-600m-int8"
    CPU_THREADS   = 4
    
    # IMPORTANT: the translation model is loaded lazily, inside get_translator().
    # RealtimeSTT starts two extra processes with the "spawn" method on macOS and
    # Windows, and each of those re-imports this file. If the model were loaded at
    # the top of the file it would load THREE times and waste about 2 GB of memory
    # on an 8 GB laptop. Do not move this out of the function.
    _TRANSLATOR = None
    _TOKENIZER = None
    
    
    def get_translator():
        global _TRANSLATOR, _TOKENIZER
        if _TRANSLATOR is None:
            print("Loading the translation model. This takes 10 to 60 seconds.")
            _TRANSLATOR = ctranslate2.Translator(
                MODEL_DIR,
                device="cpu",
                compute_type="int8",
                inter_threads=1,
                intra_threads=CPU_THREADS,
            )
            _TOKENIZER = transformers.AutoTokenizer.from_pretrained(
                MODEL_DIR, src_lang=NLLB_SOURCE
            )
        return _TRANSLATOR, _TOKENIZER
    
    
    def clear_screen():
        os.system("cls" if os.name == "nt" else "clear")
    
    
    def translate(text):
        translator, tokenizer = get_translator()
        tokens = tokenizer.convert_ids_to_tokens(tokenizer.encode(text))
        results = translator.translate_batch(
            [tokens],
            target_prefix=[[NLLB_TARGET]],
            beam_size=2,
            max_decoding_length=256,
        )
        out = results[0].hypotheses[0][1:]   # drop the language tag we forced
        return tokenizer.decode(tokenizer.convert_tokens_to_ids(out))
    
    
    def process_text(heard):
        if not heard or not heard.strip():
            return
        clear_screen()
        print("HEARD (" + LISTEN_LANGUAGE + "): " + heard)
        print("")
        print("READS (" + NLLB_TARGET + "): " + translate(heard))
        print("")
        print("-" * 60)
        print("Machine output. Check it before you use it. Press Ctrl and C to stop.")
    
    
    if __name__ == "__main__":
        get_translator()          # load once, here in the parent process only
        recorder = AudioToTextRecorder(
            model=WHISPER_MODEL,
            language=LISTEN_LANGUAGE,
            device="cpu",              # the default is "cuda"; on a laptop this MUST be "cpu"
            compute_type="int8",
            post_speech_silence_duration=0.7,
            beam_size=3,
            spinner=False,
        )
        print("Ready. Speak now.")
        try:
            while True:
                recorder.text(process_text)
        except KeyboardInterrupt:
            print("Stopping.")
        finally:
            recorder.shutdown()

    تأكد من أنها نجحت

    Back in the terminal, run: python -c "import ast; ast.parse(open('translate_live.py').read()); print('script file is valid')" It must print 'script file is valid'. If it reports a SyntaxError with a line number, the paste was incomplete or the indentation was mangled - open the file again, delete everything, and re-paste the whole block.

    ملاحظة

    Do not retype this by hand - copy and paste it. Python is sensitive to indentation and a single wrong space stops it running. The three lines under 'CHANGE THESE THREE LINES' are the only ones you ever need to edit. Do not move the model loading out of get_translator(): an earlier version of this recipe loaded it at the top of the file, and because RealtimeSTT starts two helper processes that re-import this file, the translation model was loaded three times over - about 2 GB of wasted memory on exactly the 8 GB laptops this recipe is written for, which then showed up as the machine being unbearably slow for reasons the recipe blamed on the wrong thing.

21. خطوة 21 / 23كل الأنظمة

    Set your language pair. As written, the file listens to Modern Standard Arabic and shows Kurdish Sorani. To change it, reopen the file ('notepad translate_live.py' on Windows, 'nano translate_live.py' on macOS/Ubuntu), edit only the three marked lines, and save the same way as before.

    تأكد من أنها نجحت

    After saving, run the same validity check as in step 20: python -c "import ast; ast.parse(open('translate_live.py').read()); print('script file is valid')" If you accidentally deleted a quotation mark while editing, this catches it now rather than in the middle of a meeting.

    ملاحظة

    LANGUAGES YOU CAN SPEAK (LISTEN_LANGUAGE, Whisper codes): ar = Arabic, en = English, tr = Turkish, fa = Persian, fr = French. KURDISH IS NOT ON THIS LIST AND CANNOT BE ADDED - Whisper has 100 languages and no Kurdish in any script. LANGUAGES YOU CAN READ OR WRITE (NLLB_SOURCE and NLLB_TARGET, NLLB codes): arb_Arab = Modern Standard Arabic, acm_Arab = Iraqi/Mesopotamian Arabic, apc_Arab = Levantine Arabic, arz_Arab = Egyptian Arabic, ckb_Arab = Kurdish Sorani, kmr_Latn = Kurdish Kurmanji, eng_Latn = English, tur_Latn = Turkish, pes_Arab = Persian. NLLB_SOURCE must be the same language as LISTEN_LANGUAGE - if you set LISTEN_LANGUAGE to 'en' then NLLB_SOURCE must be 'eng_Latn'. Common working pairs: Arabic in, Sorani out (ar / arb_Arab / ckb_Arab); Arabic in, English out (ar / arb_Arab / eng_Latn); English in, Sorani out (en / eng_Latn / ckb_Arab).

22. خطوة 22 / 23كل الأنظمة

    Run it. Make sure '(.venv)' is still showing at the start of your terminal line, then paste the line below.

    انسخ هذا كما هو

    python translate_live.py

    تأكد من أنها نجحت

    The first run takes 1-3 minutes to load. You should see 'Loading the translation model' ONCE - not three times - and then 'Ready. Speak now.' Speak a full sentence and pause for about a second; your words appear on the HEARD line and the translation on the READS line. If 'Loading the translation model' appears three times, the model loading was moved out of get_translator() - re-paste the script from step 20.

    ملاحظة

    Press Ctrl and C to stop. HOW LONG THE DELAY WILL BE, on a typical 4-core 8 GB laptop with no graphics card: roughly 3-8 seconds between finishing a sentence and seeing the translation. These are estimates, not benchmarks - I did not time this on real hardware. MODEL SIZE TIERS, disk / rough memory while running: tiny 75 MB / ~0.4 GB - fast and too inaccurate for Arabic, use only to check the plumbing works; base 145 MB / ~0.5 GB - still too weak for Arabic; small 484 MB / ~1.5 GB total with the translator - the recommended starting point on 8 GB; medium 1.53 GB / ~3 GB total - noticeably better on Arabic, workable on 8 GB if you close your browser, roughly twice as slow; large-v3 about 3.1 GB / ~4.5 GB total - do not attempt alongside the translator on an 8 GB laptop, it will swap to disk and crawl. On 16 GB, medium is the sweet spot. BEFORE YOU USE THIS ON ANYTHING SENSITIVE, decide where the text is going afterwards: the moment you copy a translation out of this window into Google Docs, Gmail or WhatsApp, you have uploaded it and every promise on this page stops applying.

23. خطوة 23 / 23كل الأنظمة

    Prove to yourself that nothing is leaving the laptop. Stop the program with Ctrl and C. Turn your wifi off completely - click the wifi icon and switch it off - and unplug any network cable. Then run it again and speak.

    انسخ هذا كما هو

    python translate_live.py

    تأكد من أنها نجحت

    It must reach 'Ready. Speak now.' and translate your speech exactly as before, with no internet at all. That is the proof. If instead it fails with a message mentioning 'offline', 'Connection error', 'couldn't connect to huggingface.co', or 'Could not initialize any automatic Silero VAD backend', see the note.

    ملاحظة

    This takes thirty seconds and it is the whole promise of the recipe - do it in front of your colleagues and your director rather than telling them about it. IF IT FAILS: a message about huggingface.co or 'offline' means a model file did not finish downloading, so reconnect and re-run steps 16 and 17. A message containing 'Silero VAD' means the voice-detection model is missing, which means step 14 installed the wrong option - reconnect, re-run step 14 with [recommended] exactly as written, and re-run the check in step 15 until it prints True. Do not present this tool to anyone as offline until this step passes with the wifi genuinely off.