feat: update README to reflect Apple Silicon GPU support and new features in version 3.0

Clarified installation instructions for Python 3.10 or later, specifying preferred installation method.

README.md

@@ -1,64 +1,85 @@
-## Local Transcribe with Whisper 
+## Local Transcribe with Whisper
 
-> **⚠ Note for Mac users (Apple Silicon):** This version uses `faster-whisper` (CTranslate2), which does **not** support Apple M-chip GPU acceleration. Transcription will run on CPU, which is slower than OpenAI's Whisper with Metal/CoreML support. The trade-off is a much simpler installation — no conda, no PyTorch, no admin rights. If you'd prefer M-chip GPU acceleration and don't mind a more involved setup, switch to the **classic** release:
+> **🍎 Apple Silicon GPU/NPU acceleration:** This version now supports native Apple GPU/NPU acceleration via [MLX Whisper](https://github.com/ml-explore/mlx-examples/tree/main/whisper). On Apple Silicon Macs, transcription runs on the Apple GPU and Neural Engine — no CPU fallback needed.
-> ```
-> git checkout classic
-> ```
 
-Local Transcribe with Whisper is a user-friendly desktop application that allows you to transcribe audio and video files using the Whisper ASR system, powered by [faster-whisper](https://github.com/SYSTRAN/faster-whisper) (CTranslate2). This application provides a graphical user interface (GUI) built with Python and the Tkinter library, making it easy to use even for those not familiar with programming.
+Local Transcribe with Whisper is a user-friendly desktop application that allows you to transcribe audio and video files using the Whisper ASR system, powered by [faster-whisper](https://github.com/SYSTRAN/faster-whisper) (CTranslate2) on Windows/Linux and [MLX Whisper](https://github.com/ml-explore/mlx-examples/tree/main/whisper) on Apple Silicon. This application provides a graphical user interface (GUI) built with Python and the Tkinter library, making it easy to use even for those not familiar with programming.
+
+## New in version 3.0!
+1. **Apple Silicon GPU/NPU support** — native MLX backend for Apple Silicon Macs, using Apple GPU + Neural Engine.
+2. **SRT subtitle export** — valid SubRip files alongside the existing TXT output, ready for HandBrake or any video player.
+3. **VAD filter** — removes silence, reduces hallucination, improves accuracy.
+4. **Word-level timestamps** — per-word SRT timing for precise subtitle burning.
+5. **Translation mode** — transcribe any language and translate to English in one step.
+6. **Stop button** — immediately cancel any transcription, including model downloads.
+7. **Language dropdown** — 99 languages with proper ISO codes, no more guessing formats.
+8. **Model descriptions** — speed, size, quality stars, and use case shown for every model.
 
 ## New in version 2.0!
-1. **Switched to faster-whisper** — up to 4× faster transcription with lower memory usage.
+1. **Switched to faster-whisper** — up to 4× faster transcription with lower memory usage, simpler installation.
-2. **No separate FFmpeg installation needed** — audio decoding is handled by the bundled PyAV library.
+2. **Swedish-optimised models** — [KB-Whisper](https://huggingface.co/collections/KBLab/kb-whisper) from the National Library of Sweden (KBLab)
-3. **No admin rights required** — a plain `pip install` covers everything.
+3. **No separate FFmpeg installation needed** — audio decoding is handled by the bundled PyAV library.
-4. **No PyTorch dependency** — dramatically smaller install footprint.
+4. **No admin rights required** — a plain `pip install` covers everything.
-5. **`tiny` model added** — smallest and fastest option for quick drafts.
+5. **No PyTorch dependency** — dramatically smaller install footprint.
+6. **Integrated console** - all info in the same application.
+7. **`tiny` model added** — smallest and fastest option.
+
 
 ## Features
-* Select the folder containing the audio or video files you want to transcribe. Tested with m4a video. 
+* Select the folder containing the audio or video files you want to transcribe. Tested with m4a video.
-* Choose the language of the files you are transcribing. You can either select a specific language or let the application automatically detect the language.
+* Choose the language of the files you are transcribing from a dropdown of 99 supported languages, or let the application automatically detect the language.
 * Select the Whisper model to use for the transcription. Available models include "tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large-v2", and "large-v3". Models with .en ending are better if you're transcribing English, especially the base and small models.
 * **Swedish-optimised models** — [KB-Whisper](https://huggingface.co/collections/KBLab/kb-whisper) from the National Library of Sweden (KBLab) is available in all sizes (tiny → large). These models reduce Word Error Rate by up to 47 % compared to OpenAI Whisper on Swedish speech. The language is set to Swedish automatically when a KB model is selected.
-* Enable the verbose mode to receive detailed information during the transcription process.
+* **VAD filter** — removes silence from audio before transcription, reducing hallucination and improving accuracy.
-* Monitor the progress of the transcription with the progress bar and terminal. 
+* **Word-level timestamps** — generates per-word timing in the SRT output for precise subtitle synchronization.
+* **Translation mode** — transcribes audio in any language and translates the result to English.
+* **SRT export** — valid SubRip subtitle files saved alongside TXT, ready for HandBrake or any video player.
+* Monitor the progress of the transcription with the progress bar and terminal.
 * Confirmation dialog before starting the transcription to ensure you have selected the correct folder.
 * View the transcribed text in a message box once the transcription is completed.
+* **Stop button** — immediately cancel transcription, including model downloads.
 
 ## Installation
 ### Get the files
-Download the zip folder and extract it to your preferred working folder.  
+Download the zip folder and extract it to your preferred working folder.
-![](images/Picture1.png)  
+![](images/Picture1.png)
 Or by cloning the repository with:
 ```
-git clone https://github.com/soderstromkr/transcribe.git
+git clone https://gitea.kobim.cloud/kobim/whisper-local-transcribe.git
 ```
-### Python Version **(any platform including Mac users)**
+### Prerequisites
-1. Install Python 3.10 or later. You can download it from [python.org](https://www.python.org/downloads/). During installation, **check "Add Python to PATH"**. No administrator rights are needed if you install for your user only.
+Install **Python 3.10 or later**. Some IT policies allow installing from the Microsoft Store or Mac equivalent. However, I would prefer an install from [python.org](https://www.python.org/downloads/). During installation, **check "Add Python to PATH"**. No administrator rights are needed if you install for your user only.
 
-2. Run the installer. Open a terminal (Command Prompt on Windows, Terminal on Mac/Linux) in the project folder and run:
+### Run on Windows
+Double-click `run_Windows.bat` — it will auto-install everything on first run.
+
+### Run on Mac / Linux
+Run `./run_Mac.sh` — it will auto-install everything on first run. See [Mac instructions](Mac_instructions.md) for details.
+
+> **Note:** The first run with a given model will download it (~75 MB for base, ~500 MB for medium). After that, everything works offline.
+
+### Manual installation (if the launchers don't work)
+If `run_Windows.bat` or `run_Mac.sh` fails (e.g. Python isn't on PATH, or permissions issues), open a terminal in the project folder and run these steps manually:
+```
+python -m venv .venv
+```
+Activate the virtual environment:
+- **Windows:** `.venv\Scripts\activate`
+- **Mac / Linux:** `source .venv/bin/activate`
+
+Then install and run:
 ```
 python install.py
-```
-This will:
-- Install all required packages (including bundled FFmpeg — no separate install needed)
-- **Auto-detect your NVIDIA GPU** and ask if you want GPU acceleration
-- No conda, no admin rights required
-
-Alternatively, you can install manually with `pip install -r requirements.txt`.
-
-3. Run the app: 
-    1. For **Windows**: double-click `run_Windows.bat` (it will auto-install on first run) or run:
-```
 python app.py
 ```
-    2. For **Mac / Linux**: run `./run_Mac.sh` (auto-installs on first run). See [Mac instructions](Mac_instructions.md) for details.
-
-    **Note** The first run with a given model will download it (~75 MB for base, ~500 MB for medium). After that, everything works offline.
 
 ## GPU Support
+### Apple Silicon
+On Macs with Apple Silicon, the app automatically uses the **MLX backend**, which runs inference on the Apple GPU and Neural Engine. No additional setup is needed — just install and run. MLX models are downloaded from HuggingFace on first use.
+
+### NVIDIA GPUs
 This program **does support running on NVIDIA GPUs**, which can significantly speed up transcription times. faster-whisper uses CTranslate2, which requires NVIDIA CUDA libraries for GPU acceleration.
 
-### Automatic Detection
+#### Automatic Detection
 The `install.py` script **automatically detects NVIDIA GPUs** and will ask if you want to install GPU support. If you skipped it during installation, you can add it anytime:
 ```
 pip install nvidia-cublas-cu12 nvidia-cudnn-cu12
@@ -66,7 +87,7 @@ pip install nvidia-cublas-cu12 nvidia-cudnn-cu12
 
 **Note:** Make sure your NVIDIA GPU drivers are up to date. You can check by running `nvidia-smi` in your terminal. The program will automatically detect and use your GPU if available, otherwise it falls back to CPU.
 
-### Verifying GPU Support
+#### Verifying GPU Support
 After installation, you can verify that your GPU is available by running:
 ```python
 import ctranslate2
@@ -75,14 +96,16 @@ print(ctranslate2.get_supported_compute_types("cuda"))
 If this returns a list containing `"float16"`, GPU acceleration is working.
 
 ## Usage
-1. Launch the app — the built-in console panel at the bottom shows a welcome message and all progress updates.
+1. Launch the app — the built-in console panel at the bottom shows a welcome message and all progress updates. The backend indicator at the bottom shows which inference engine is active (MLX · Apple GPU/NPU, CUDA · GPU, or CPU · int8).
 2. Select the folder containing the audio or video files you want to transcribe by clicking the "Browse" button next to the "Folder" label. This will open a file dialog where you can navigate to the desired folder. Remember, you won't be choosing individual files but whole folders!
-3. Enter the desired language for the transcription in the "Language" field. You can either select a language or leave it blank to enable automatic language detection.
+3. Select the language from the dropdown — 99 languages are available, or leave it on "Auto-detect". For English-only models (.en) the language is locked to English; for KB Swedish models it's locked to Swedish.
-4. Choose the Whisper model to use for the transcription from the dropdown list next to the "Model" label.
+4. Choose the Whisper model to use for the transcription from the dropdown list next to the "Model" label. A description below shows speed, size, quality stars, and recommended use case for each model.
-5. Click the "Transcribe" button to start the transcription. The button will be disabled during the process to prevent multiple transcriptions at once.
+5. Toggle advanced options if needed: **VAD filter**, **Word-level timestamps**, or **Translate to English**.
-6. Monitor progress in the embedded console panel — it shows model loading, per-file progress, and segment timestamps in real time.
+6. Click the "Transcribe" button to start the transcription. Use the "Stop" button to cancel at any time.
-7. Once the transcription is completed, a message box will appear displaying the result. Click "OK" to close it.
+7. Monitor progress in the embedded console panel — it shows model loading, per-file progress, and segment timestamps in real time.
-8. You can run the application again or quit at any time by clicking the "Quit" button.
+8. Once the transcription is completed, a message box will appear displaying the result. Click "OK" to close it.
+9. Transcriptions are saved as both `.txt` (human-readable) and `.srt` (SubRip subtitles) in the `transcriptions/` folder within the selected directory.
+10. You can run the application again or quit at any time by clicking the "Quit" button.
 
 ## Jupyter Notebook
 Don't want fancy EXEs or GUIs? Use the function as is. See [example](example.ipynb) for an implementation on Jupyter Notebook.

app.py

@@ -4,7 +4,8 @@ import tkinter as tk
 from tkinter import ttk
 from tkinter import filedialog
 from tkinter import messagebox
-from src._LocalTranscribe import transcribe, get_path
+from src._LocalTranscribe import transcribe, get_path, detect_backend, _transcribe_worker_process
+import multiprocessing as mp
 import customtkinter
 import threading
 
@@ -46,11 +47,121 @@ HF_MODEL_MAP = {
     'KB Swedish (large)':  'KBLab/kb-whisper-large',
 }
 
+# Per-model info shown in the UI description label
+# (speed, size, quality stars, suggested use)
+MODEL_INFO = {
+    'tiny':                 ('Very fast',   '~75 MB',   '★★☆☆☆', 'Quick drafts & testing'),
+    'tiny.en':              ('Very fast',   '~75 MB',   '★★☆☆☆', 'Quick drafts & testing (English only)'),
+    'base':                 ('Fast',        '~145 MB',  '★★★☆☆', 'Notes & short podcasts'),
+    'base.en':              ('Fast',        '~145 MB',  '★★★☆☆', 'Notes & short podcasts (English only)'),
+    'small':                ('Balanced',    '~485 MB',  '★★★★☆', 'Everyday use'),
+    'small.en':             ('Balanced',    '~485 MB',  '★★★★☆', 'Everyday use (English only)'),
+    'medium':               ('Accurate',    '~1.5 GB',  '★★★★☆', 'Professional content'),
+    'medium.en':            ('Accurate',    '~1.5 GB',  '★★★★☆', 'Professional content (English only)'),
+    'large-v2':             ('Slow',        '~3 GB',    '★★★★★', 'Maximum accuracy'),
+    'large-v3':             ('Slow',        '~3 GB',    '★★★★★', 'Maximum accuracy (recommended)'),
+    'KB Swedish (tiny)':    ('Very fast',   '~75 MB',   '★★★☆☆', 'Swedish — optimised by KBLab'),
+    'KB Swedish (base)':    ('Fast',        '~145 MB',  '★★★☆☆', 'Swedish — optimised by KBLab'),
+    'KB Swedish (small)':   ('Balanced',    '~485 MB',  '★★★★☆', 'Swedish — optimised by KBLab'),
+    'KB Swedish (medium)':  ('Accurate',    '~1.5 GB',  '★★★★☆', 'Swedish — optimised by KBLab'),
+    'KB Swedish (large)':   ('Slow',        '~3 GB',    '★★★★★', 'Swedish — KBLab, best accuracy'),
+}
+
 
 
 customtkinter.set_appearance_mode("System")
 customtkinter.set_default_color_theme("blue")  # Themes: blue (default), dark-blue, green
-firstclick = True
+
+# All languages supported by Whisper  (display label → ISO code; None = auto-detect)
+WHISPER_LANGUAGES = {
+    'Auto-detect':          None,
+    'Afrikaans (af)':       'af',   'Albanian (sq)':        'sq',
+    'Amharic (am)':         'am',   'Arabic (ar)':          'ar',
+    'Armenian (hy)':        'hy',   'Assamese (as)':        'as',
+    'Azerbaijani (az)':     'az',   'Bashkir (ba)':         'ba',
+    'Basque (eu)':          'eu',   'Belarusian (be)':      'be',
+    'Bengali (bn)':         'bn',   'Bosnian (bs)':         'bs',
+    'Breton (br)':          'br',   'Bulgarian (bg)':       'bg',
+    'Catalan (ca)':         'ca',   'Chinese (zh)':         'zh',
+    'Croatian (hr)':        'hr',   'Czech (cs)':           'cs',
+    'Danish (da)':          'da',   'Dutch (nl)':           'nl',
+    'English (en)':         'en',   'Estonian (et)':        'et',
+    'Faroese (fo)':         'fo',   'Finnish (fi)':         'fi',
+    'French (fr)':          'fr',   'Galician (gl)':        'gl',
+    'Georgian (ka)':        'ka',   'German (de)':          'de',
+    'Greek (el)':           'el',   'Gujarati (gu)':        'gu',
+    'Haitian Creole (ht)':  'ht',   'Hausa (ha)':           'ha',
+    'Hawaiian (haw)':       'haw',  'Hebrew (he)':          'he',
+    'Hindi (hi)':           'hi',   'Hungarian (hu)':       'hu',
+    'Icelandic (is)':       'is',   'Indonesian (id)':      'id',
+    'Italian (it)':         'it',   'Japanese (ja)':        'ja',
+    'Javanese (jw)':        'jw',   'Kannada (kn)':         'kn',
+    'Kazakh (kk)':          'kk',   'Khmer (km)':           'km',
+    'Korean (ko)':          'ko',   'Lao (lo)':             'lo',
+    'Latin (la)':           'la',   'Latvian (lv)':         'lv',
+    'Lingala (ln)':         'ln',   'Lithuanian (lt)':      'lt',
+    'Luxembourgish (lb)':   'lb',   'Macedonian (mk)':      'mk',
+    'Malagasy (mg)':        'mg',   'Malay (ms)':           'ms',
+    'Malayalam (ml)':       'ml',   'Maltese (mt)':         'mt',
+    'Maori (mi)':           'mi',   'Marathi (mr)':         'mr',
+    'Mongolian (mn)':       'mn',   'Myanmar (my)':         'my',
+    'Nepali (ne)':          'ne',   'Norwegian (no)':       'no',
+    'Occitan (oc)':         'oc',   'Pashto (ps)':          'ps',
+    'Persian (fa)':         'fa',   'Polish (pl)':          'pl',
+    'Portuguese (pt)':      'pt',   'Punjabi (pa)':         'pa',
+    'Romanian (ro)':        'ro',   'Russian (ru)':         'ru',
+    'Sanskrit (sa)':        'sa',   'Serbian (sr)':         'sr',
+    'Shona (sn)':           'sn',   'Sindhi (sd)':          'sd',
+    'Sinhala (si)':         'si',   'Slovak (sk)':          'sk',
+    'Slovenian (sl)':       'sl',   'Somali (so)':          'so',
+    'Spanish (es)':         'es',   'Sundanese (su)':       'su',
+    'Swahili (sw)':         'sw',   'Swedish (sv)':         'sv',
+    'Tagalog (tl)':         'tl',   'Tajik (tg)':           'tg',
+    'Tamil (ta)':           'ta',   'Tatar (tt)':           'tt',
+    'Telugu (te)':          'te',   'Thai (th)':            'th',
+    'Tibetan (bo)':         'bo',   'Turkish (tr)':         'tr',
+    'Turkmen (tk)':         'tk',   'Ukrainian (uk)':       'uk',
+    'Urdu (ur)':            'ur',   'Uzbek (uz)':           'uz',
+    'Vietnamese (vi)':      'vi',   'Welsh (cy)':           'cy',
+    'Yiddish (yi)':         'yi',   'Yoruba (yo)':          'yo',
+}
+
+
+def _language_options_for_model(model_name):
+    """Return (values, default, state) for the language combobox given a model name."""
+    if model_name.endswith('.en'):
+        return ['English (en)'], 'English (en)', 'disabled'
+    if model_name.startswith('KB Swedish'):
+        return ['Swedish (sv)'], 'Swedish (sv)', 'disabled'
+    return list(WHISPER_LANGUAGES.keys()), 'Auto-detect', 'readonly'
+
+
+def _set_app_icon(root):
+    """Set app icon when supported, without crashing on unsupported platforms."""
+    base_dir = os.path.dirname(os.path.abspath(__file__))
+    icon_path = os.path.join(base_dir, "images", "icon.ico")
+
+    if not os.path.exists(icon_path):
+        return
+
+    try:
+        root.iconbitmap(icon_path)
+    except tk.TclError:
+        # Some Linux Tk builds don't accept .ico for iconbitmap.
+        pass
+
+
+def _apply_display_scaling(root):
+    """Auto-scale UI for high-resolution displays (e.g., 4K)."""
+    try:
+        screen_w = root.winfo_screenwidth()
+        screen_h = root.winfo_screenheight()
+        scale = min(screen_w / 1920.0, screen_h / 1080.0)
+        scale = max(1.0, min(scale, 2.0))
+        customtkinter.set_widget_scaling(scale)
+        customtkinter.set_window_scaling(scale)
+    except Exception:
+        pass
 
 class App:
     def __init__(self, master):
@@ -66,22 +177,16 @@ class App:
         self.path_entry.insert(0, os.path.join(os.getcwd(), 'sample_audio'))
         self.path_entry.pack(side=tk.LEFT, fill=tk.X, expand=True)
         customtkinter.CTkButton(path_frame, text="Browse", command=self.browse, font=font).pack(side=tk.LEFT, padx=5)
-        # Language frame        
+        # Language frame
-        #thanks to pommicket from Stackoverflow for this fix
-        def on_entry_click(event):
-            """function that gets called whenever entry is clicked"""        
-            global firstclick
-            if firstclick: # if this is the first time they clicked it
-                firstclick = False
-                self.language_entry.delete(0, "end") # delete all the text in the entry
         language_frame = customtkinter.CTkFrame(master)
         language_frame.pack(fill=tk.BOTH, padx=10, pady=10)
         customtkinter.CTkLabel(language_frame, text="Language:", font=font).pack(side=tk.LEFT, padx=5)
-        self.language_entry = customtkinter.CTkEntry(language_frame, width=50, font=('Roboto', 12, 'italic'))
+        _lang_values, _lang_default, _lang_state = _language_options_for_model('medium')
-        self.default_language_text = "Enter language (or ignore to auto-detect)"
+        self.language_combobox = customtkinter.CTkComboBox(
-        self.language_entry.insert(0, self.default_language_text)
+            language_frame, width=50, state=_lang_state,
-        self.language_entry.bind('<FocusIn>', on_entry_click)
+            values=_lang_values, font=font_b)
-        self.language_entry.pack(side=tk.LEFT, fill=tk.X, expand=True)
+        self.language_combobox.set(_lang_default)
+        self.language_combobox.pack(side=tk.LEFT, fill=tk.X, expand=True)
         # Model frame
         models = ['tiny', 'tiny.en', 'base', 'base.en',
                   'small', 'small.en', 'medium', 'medium.en',
@@ -96,16 +201,54 @@ class App:
         # ComboBox frame
         self.model_combobox = customtkinter.CTkComboBox(
             model_frame, width=50, state="readonly",
-            values=models, font=font_b)
+            values=models, font=font_b,
+            command=self._on_model_change)
         self.model_combobox.set('medium')  # Set the default value
         self.model_combobox.pack(side=tk.LEFT, fill=tk.X, expand=True)
+        # Model description label
+        self.model_desc_label = customtkinter.CTkLabel(
+            master, text=self._model_desc_text('medium'),
+            font=('Roboto', 11), text_color=('#555555', '#aaaaaa'),
+            anchor='w')
+        self.model_desc_label.pack(fill=tk.X, padx=14, pady=(0, 4))
+        # Timestamps toggle
+        ts_frame = customtkinter.CTkFrame(master)
+        ts_frame.pack(fill=tk.BOTH, padx=10, pady=10)
+        self.timestamps_var = tk.BooleanVar(value=True)
+        self.timestamps_switch = customtkinter.CTkSwitch(
+            ts_frame, text="Include timestamps in transcription",
+            variable=self.timestamps_var, font=font_b)
+        self.timestamps_switch.pack(side=tk.LEFT, padx=5)
+        # Advanced options frame
+        adv_frame = customtkinter.CTkFrame(master)
+        adv_frame.pack(fill=tk.BOTH, padx=10, pady=10)
+        self.vad_var = tk.BooleanVar(value=False)
+        customtkinter.CTkSwitch(
+            adv_frame, text="VAD filter (remove silence)",
+            variable=self.vad_var, font=font_b).pack(side=tk.LEFT, padx=5)
+        self.word_ts_var = tk.BooleanVar(value=False)
+        customtkinter.CTkSwitch(
+            adv_frame, text="Word-level timestamps",
+            variable=self.word_ts_var, font=font_b).pack(side=tk.LEFT, padx=5)
+        self.translate_var = tk.BooleanVar(value=False)
+        customtkinter.CTkSwitch(
+            adv_frame, text="Translate to English",
+            variable=self.translate_var, font=font_b).pack(side=tk.LEFT, padx=5)
         # Progress Bar
         self.progress_bar = ttk.Progressbar(master, length=200, mode='indeterminate')
+        # Worker process handle (replaces thread+stop_event for true immediate cancellation)
+        self._proc = None
+        self._parent_conn = None
+        self._child_conn = None
         # Button actions frame
         button_frame = customtkinter.CTkFrame(master)
         button_frame.pack(fill=tk.BOTH, padx=10, pady=10)
         self.transcribe_button = customtkinter.CTkButton(button_frame, text="Transcribe", command=self.start_transcription, font=font)
         self.transcribe_button.pack(side=tk.LEFT, padx=5, pady=10, fill=tk.X, expand=True)
+        self.stop_button = customtkinter.CTkButton(
+            button_frame, text="Stop", command=self._stop_transcription, font=font,
+            fg_color="#c0392b", hover_color="#922b21", state=tk.DISABLED)
+        self.stop_button.pack(side=tk.LEFT, padx=5, pady=10, fill=tk.X, expand=True)
         customtkinter.CTkButton(button_frame, text="Quit", command=master.quit, font=font).pack(side=tk.RIGHT, padx=5, pady=10, fill=tk.X, expand=True)
 
         # ── Embedded console / log panel ──────────────────────────────────
@@ -120,11 +263,59 @@ class App:
         sys.stdout = _ConsoleRedirector(self.log_box)
         sys.stderr = _ConsoleRedirector(self.log_box)
 
+        # Backend indicator
+        _bi = detect_backend()
+        backend_label = customtkinter.CTkLabel(
+            master,
+            text=f"Backend: {_bi['label']}",
+            font=('Roboto', 11),
+            text_color=("#555555", "#aaaaaa"),
+            anchor='e',
+        )
+        backend_label.pack(fill=tk.X, padx=12, pady=(0, 2))
+
         # Welcome message (shown after redirect so it appears in the panel)
         print("Welcome to Local Transcribe with Whisper! \U0001f600")
         print("Transcriptions will be saved automatically.")
         print("─" * 46)
     # Helper functions
+    def _stop_transcription(self):
+        self.stop_button.configure(state=tk.DISABLED)
+        if self._proc and self._proc.is_alive():
+            self._proc.terminate()
+            try:
+                self._proc.join(timeout=3)
+            except Exception:
+                pass
+            if self._proc.is_alive():
+                self._proc.kill()
+                try:
+                    self._proc.join(timeout=1)
+                except Exception:
+                    pass
+        # Close pipe ends — no semaphores, so no leak
+        for conn in (self._parent_conn, self._child_conn):
+            try:
+                if conn:
+                    conn.close()
+            except Exception:
+                pass
+        self._parent_conn = self._child_conn = None
+        print("⛔  Transcription stopped by user.")
+
+    def _model_desc_text(self, model_name):
+        info = MODEL_INFO.get(model_name)
+        if not info:
+            return ''
+        speed, size, stars, use = info
+        return f'{stars}  {speed}  ·  {size}  ·  {use}'
+
+    def _on_model_change(self, selected):
+        self.model_desc_label.configure(text=self._model_desc_text(selected))
+        values, default, state = _language_options_for_model(selected)
+        self.language_combobox.configure(values=values, state=state)
+        self.language_combobox.set(default)
+
     # Browsing
     def browse(self):
         initial_dir = os.getcwd()
@@ -133,64 +324,83 @@ class App:
         self.path_entry.insert(0, folder_path)
     # Start transcription
     def start_transcription(self):
-        # Disable transcribe button
-        self.transcribe_button.configure(state=tk.DISABLED)
-        # Start a new thread for the transcription process
-        threading.Thread(target=self.transcribe_thread).start()
-    # Threading
-    def transcribe_thread(self):
-        path = self.path_entry.get()
         model_display = self.model_combobox.get()
-        # Ignore the visual separator
         if model_display.startswith('─'):
             messagebox.showinfo("Invalid selection", "Please select a model, not the separator line.")
-            self.transcribe_button.configure(state=tk.NORMAL)
             return
+        self.transcribe_button.configure(state=tk.DISABLED)
+        self.stop_button.configure(state=tk.NORMAL)
+        path = self.path_entry.get()
         model = HF_MODEL_MAP.get(model_display, model_display)
-        language = self.language_entry.get()
+        lang_label = self.language_combobox.get()
-        # Auto-set Swedish for KB models
+        language = WHISPER_LANGUAGES.get(lang_label, lang_label) if lang_label else None
-        is_kb_model = model_display.startswith('KB Swedish')
+        timestamps = self.timestamps_var.get()
-        # Check if the language field has the default text or is empty
+        vad_filter = self.vad_var.get()
-        if is_kb_model:
+        word_timestamps = self.word_ts_var.get()
-            language = 'sv'
+        translate = self.translate_var.get()
-        elif language == self.default_language_text or not language.strip():
+        glob_file = get_path(path)
-            language = None  # This is the same as passing nothing
-        verbose = True   # always show transcription progress in the console panel
-        # Show progress bar
         self.progress_bar.pack(fill=tk.X, padx=5, pady=5)
         self.progress_bar.start()
-        # Setting path and files
+        self._parent_conn, self._child_conn = mp.Pipe(duplex=False)
-        glob_file = get_path(path)
+        self._proc = mp.Process(
-        #messagebox.showinfo("Message", "Starting transcription!")
+            target=_transcribe_worker_process,
-        # Start transcription
+            args=(self._child_conn, path, glob_file, model, language, True, timestamps),
+            kwargs={"vad_filter": vad_filter, "word_timestamps": word_timestamps, "translate": translate},
+            daemon=True,
+        )
+        self._proc.start()
+        self._child_conn.close()  # parent doesn't write; close its write-end
+        self._child_conn = None
+        self.master.after(100, self._poll_worker)
+
+    def _poll_worker(self):
+        done = False
+        result = None
         try:
-            output_text = transcribe(path, glob_file, model, language, verbose)
+            while self._parent_conn and self._parent_conn.poll():
-        except UnboundLocalError:
+                msg = self._parent_conn.recv()
-            messagebox.showinfo("Files not found error!", 'Nothing found, choose another folder.')
+                if isinstance(msg, tuple) and msg[0] == '__done__':
+                    done = True
+                    result = msg[1]
+                else:
+                    sys.stdout.write(msg)
+                    sys.stdout.flush()
+        except EOFError:
+            # Child closed the pipe (normal completion or kill)
+            done = True
+        except Exception:
             pass
-        except ValueError:
+        if done or (self._proc and not self._proc.is_alive()):
-            messagebox.showinfo("Invalid language name, you might have to clear the default text to continue!")
+            if self._parent_conn:
-        # Hide progress bar
+                try:
+                    self._parent_conn.close()
+                except Exception:
+                    pass
+                self._parent_conn = None
+            self._on_transcription_done(result)
+        else:
+            self.master.after(100, self._poll_worker)
+
+    def _on_transcription_done(self, output_text):
         self.progress_bar.stop()
         self.progress_bar.pack_forget()
-        # Enable transcribe button
+        self.stop_button.configure(state=tk.DISABLED)
         self.transcribe_button.configure(state=tk.NORMAL)
-        # Recover output text
+        if output_text:
-        try:
+            title = "Finished!" if not output_text.startswith('⚠') else "Error"
-            messagebox.showinfo("Finished!", output_text)
+            messagebox.showinfo(title, output_text)
-        except UnboundLocalError:
-            pass
 
 if __name__ == "__main__":
     # Setting custom themes
     root = customtkinter.CTk()
+    _apply_display_scaling(root)
     root.title("Local Transcribe with Whisper")
     # Geometry — taller to accommodate the embedded console panel
     width, height = 550, 560
     root.geometry('{}x{}'.format(width, height))
     root.minsize(450, 480)
-    # Icon 
+    # Icon (best-effort; ignored on platforms/builds without .ico support)
-    root.iconbitmap('images/icon.ico')
+    _set_app_icon(root)
     # Run
     app = App(root)
     root.mainloop()

requirements.txt

@@ -1,2 +1,3 @@
 faster-whisper
+mlx-whisper
 customtkinter

sample_audio/transcriptions/Armstrong_Small_Step.txt

@@ -1,2 +1,4 @@
 Armstrong_Small_Step
-[0:00:00 --> 0:00:07]: That's one small step for man, one giant leap for mankind.
+────────────────────────────────────────
+
+That's one small step for man, one giant leap for mankind.

sample_audio/transcriptions/Axel_Pettersson_röstinspelning.txt

@@ -1,2 +1,4 @@
 Axel_Pettersson_röstinspelning
-[0:00:00 --> 0:00:15]: Hej, jag heter Axel Pettersson, jag föddes i Örebro 1976. Jag har varit Wikipedia sen 2008 och jag har översatt röstintroduktionsprojektet till svenska.
+────────────────────────────────────────
+
+Hej, jag heter Axel Pettersson, jag föddes i Örebro 1976. Jag har varit Wikipedia sen 2008 och jag har översatt röstintroduktionsprojektet till svenska.

src/_LocalTranscribe.py

@@ -1,65 +1,210 @@
 import os
 import sys
+import platform
 import datetime
+import time
 import site
 from glob import glob
 
 # ---------------------------------------------------------------------------
 # CUDA setup — must happen before importing faster_whisper / ctranslate2
 # ---------------------------------------------------------------------------
-def _setup_cuda_dlls():
+def _setup_cuda_libs():
-    """Add NVIDIA pip-package DLL dirs to the DLL search path (Windows only).
+    """Register NVIDIA pip-package lib dirs so ctranslate2 finds CUDA at runtime.
 
-    pip-installed nvidia-cublas-cu12 / nvidia-cudnn-cu12 place their .dll
+    pip-installed nvidia-cublas-cu12 / nvidia-cudnn-cu12 place their shared
-    files inside the site-packages tree.  Python 3.8+ on Windows does NOT
+    libraries inside the site-packages tree.  Neither Windows nor Linux
-    search PATH for DLLs loaded via ctypes/LoadLibrary, so we must
+    automatically search those directories, so we must register them
-    explicitly register every nvidia/*/bin and nvidia/*/lib directory using
+    explicitly:
-    os.add_dll_directory *and* prepend them to PATH (some native extensions
+      - Windows: os.add_dll_directory() + PATH
-    still rely on PATH).
+      - Linux:   LD_LIBRARY_PATH  (read by the dynamic linker)
     """
-    if sys.platform != "win32":
-        return
     try:
-        for sp in site.getsitepackages():
+        sp_dirs = site.getsitepackages()
-            nvidia_root = os.path.join(sp, "nvidia")
+    except AttributeError:
-            if not os.path.isdir(nvidia_root):
+        # virtualenv without site-packages helper
-                continue
+        sp_dirs = [os.path.join(sys.prefix, "lib",
-            for pkg in os.listdir(nvidia_root):
+                                "python" + ".".join(map(str, sys.version_info[:2])),
-                for sub in ("bin", "lib"):
+                                "site-packages")]
-                    d = os.path.join(nvidia_root, pkg, sub)
-                    if os.path.isdir(d):
-                        os.environ["PATH"] = d + os.pathsep + os.environ.get("PATH", "")
-                        try:
-                            os.add_dll_directory(d)
-                        except (OSError, AttributeError):
-                            pass
-    except Exception:
-        pass
 
-_setup_cuda_dlls()
+    for sp in sp_dirs:
+        nvidia_root = os.path.join(sp, "nvidia")
+        if not os.path.isdir(nvidia_root):
+            continue
+        for pkg in os.listdir(nvidia_root):
+            for sub in ("bin", "lib"):
+                d = os.path.join(nvidia_root, pkg, sub)
+                if not os.path.isdir(d):
+                    continue
+                if sys.platform == "win32":
+                    os.environ["PATH"] = d + os.pathsep + os.environ.get("PATH", "")
+                    try:
+                        os.add_dll_directory(d)
+                    except (OSError, AttributeError):
+                        pass
+                else:
+                    # Linux / macOS — prepend to LD_LIBRARY_PATH
+                    ld = os.environ.get("LD_LIBRARY_PATH", "")
+                    if d not in ld:
+                        os.environ["LD_LIBRARY_PATH"] = d + (":" + ld if ld else "")
+                        # Also load via ctypes so already-started process sees it
+                        import ctypes
+                        try:
+                            for so in sorted(os.listdir(d)):
+                                if so.endswith(".so") or ".so." in so:
+                                    ctypes.cdll.LoadLibrary(os.path.join(d, so))
+                        except OSError:
+                            pass
+
+_setup_cuda_libs()
 
 from faster_whisper import WhisperModel
 
 
-def _detect_device():
+SUPPORTED_EXTENSIONS = {
-    """Return (device, compute_type) for the best available backend."""
+    ".wav", ".mp3", ".m4a", ".flac", ".ogg", ".wma", ".aac",
+    ".mp4", ".mkv", ".mov", ".webm", ".avi", ".mpeg", ".mpg",
+}
+
+
+# ---------------------------------------------------------------------------
+# MLX model map  (Apple Silicon only)
+# ---------------------------------------------------------------------------
+_MLX_MODEL_MAP = {
+    "tiny":     "mlx-community/whisper-tiny-mlx",
+    "base":     "mlx-community/whisper-base-mlx",
+    "small":    "mlx-community/whisper-small-mlx",
+    "medium":   "mlx-community/whisper-medium-mlx",
+    "large-v2": "mlx-community/whisper-large-v2-mlx",
+    "large-v3": "mlx-community/whisper-large-v3-mlx",
+}
+
+
+def detect_backend():
+    """Return the best available inference backend.
+
+    Returns a dict with keys:
+        backend      : "mlx" | "cuda" | "cpu"
+        device       : device string for WhisperModel (cuda / cpu)
+        compute_type : compute type string for WhisperModel
+        label        : human-readable label for UI display
+    """
+    # Apple Silicon → try MLX (GPU + Neural Engine via Apple MLX)
+    if sys.platform == "darwin" and platform.machine() == "arm64":
+        try:
+            import mlx_whisper  # noqa: F401
+            return {
+                "backend": "mlx",
+                "device": "cpu",
+                "compute_type": "int8",
+                "label": "MLX · Apple GPU/NPU",
+            }
+        except ImportError:
+            pass
+
+    # NVIDIA CUDA
     try:
         import ctranslate2
         cuda_types = ctranslate2.get_supported_compute_types("cuda")
         if "float16" in cuda_types:
-            return "cuda", "float16"
+            return {
+                "backend": "cuda",
+                "device": "cuda",
+                "compute_type": "float16",
+                "label": "CUDA · GPU",
+            }
     except Exception:
         pass
-    return "cpu", "int8"
+
+    return {
+        "backend": "cpu",
+        "device": "cpu",
+        "compute_type": "int8",
+        "label": "CPU · int8",
+    }
+
+
+def _decode_audio_pyav(file_path):
+    """Decode any audio/video file to a float32 mono 16 kHz numpy array.
+
+    Uses PyAV (bundled FFmpeg) — no external ffmpeg binary required.
+    Returns (audio_array, duration_seconds).
+    """
+    import av
+    import numpy as np
+
+    with av.open(file_path) as container:
+        duration = float(container.duration) / 1_000_000  # microseconds → seconds
+        stream = container.streams.audio[0]
+        resampler = av.AudioResampler(format="fltp", layout="mono", rate=16000)
+        chunks = []
+        for frame in container.decode(stream):
+            for out in resampler.resample(frame):
+                if out:
+                    chunks.append(out.to_ndarray()[0])
+        # Flush resampler
+        for out in resampler.resample(None):
+            if out:
+                chunks.append(out.to_ndarray()[0])
+
+    if not chunks:
+        return np.zeros(0, dtype=np.float32), duration
+    return np.concatenate(chunks, axis=0), duration
+
+
+def _transcribe_mlx_file(file, mlx_model_id, language, timestamps, verbose, vad_filter=False, word_timestamps=False, translate=False):
+    """Transcribe a single file with mlx-whisper (Apple GPU/NPU).
+
+    Decodes audio via PyAV (no system ffmpeg needed), then runs MLX inference.
+    Returns (segments_as_dicts, audio_duration_seconds).
+    Segments have dict keys: 'start', 'end', 'text'.
+    """
+    import mlx_whisper
+
+    audio_array, duration = _decode_audio_pyav(file)
+
+    decode_opts = {}
+    if language:
+        decode_opts["language"] = language
+    if translate:
+        decode_opts["task"] = "translate"
+    if word_timestamps:
+        decode_opts["word_timestamps"] = True
+
+    result = mlx_whisper.transcribe(
+        audio_array,
+        path_or_hf_repo=mlx_model_id,
+        verbose=(True if verbose else None),
+        **decode_opts,
+    )
+    segments = result["segments"]
+    audio_duration = segments[-1]["end"] if segments else duration
+    return segments, audio_duration
+
+
+def _srt_timestamp(seconds):
+    """Convert seconds (float) to SRT timestamp format HH:MM:SS,mmm."""
+    ms = round(seconds * 1000)
+    h, ms = divmod(ms, 3_600_000)
+    m, ms = divmod(ms, 60_000)
+    s, ms = divmod(ms, 1000)
+    return f"{h:02d}:{m:02d}:{s:02d},{ms:03d}"
 
 
 # Get the path
 def get_path(path):
-    glob_file = glob(path + '/*')
+    all_items = glob(path + '/*')
-    return glob_file
+    media_files = []
+    for item in all_items:
+        if not os.path.isfile(item):
+            continue
+        _, ext = os.path.splitext(item)
+        if ext.lower() in SUPPORTED_EXTENSIONS:
+            media_files.append(item)
+    return sorted(media_files)
 
 # Main function
-def transcribe(path, glob_file, model=None, language=None, verbose=False):
+def transcribe(path, glob_file, model=None, language=None, verbose=False, timestamps=True, stop_event=None, vad_filter=False, word_timestamps=False, translate=False):
     """
     Transcribes audio files in a specified folder using faster-whisper (CTranslate2).
 
@@ -90,53 +235,204 @@ def transcribe(path, glob_file, model=None, language=None, verbose=False):
     SEP = "─" * 46
 
     # ── Step 1: Detect hardware ──────────────────────────────────────
-    device, compute_type = _detect_device()
+    backend_info = detect_backend()
-    print(f"⚙  Device: {device}  |  Compute: {compute_type}")
+    backend   = backend_info["backend"]
+    device    = backend_info["device"]
+    compute_type = backend_info["compute_type"]
+    print(f"⚙  Backend: {backend_info['label']}")
 
-    # ── Step 2: Load model ───────────────────────────────────────────
+    # ── Step 1b: MLX path (Apple GPU/NPU) ───────────────────────────
+    if backend == "mlx":
+        mlx_model_id = _MLX_MODEL_MAP.get(model)
+        if mlx_model_id is None:
+            print(f"⚠  Model '{model}' is not available in MLX format.")
+            print("   Falling back to faster-whisper on CPU (int8).")
+            backend = "cpu"
+            device, compute_type = "cpu", "int8"
+        else:
+            # ── Step 2 (MLX): load + transcribe ─────────────────────
+            print(f"⏳ Loading MLX model '{model}' — downloading if needed...")
+            print("✅ Model ready!")
+            print(SEP)
+
+            total_files = len(glob_file)
+            print(f"📂 Found {total_files} supported media file(s) in folder")
+            print(SEP)
+
+            if total_files == 0:
+                output_text = '⚠  No supported media files found — try another folder.'
+                print(output_text)
+                print(SEP)
+                return output_text
+
+            files_transcripted = []
+            file_num = 0
+            for file in glob_file:
+                if stop_event and stop_event.is_set():
+                    print("⛔  Transcription stopped by user.")
+                    break
+                title = os.path.basename(file).split('.')[0]
+                file_num += 1
+                print(f"\n{'─' * 46}")
+                print(f"📄 File {file_num}/{total_files}: {title}")
+                try:
+                    t_start = time.time()
+                    segments, audio_duration = _transcribe_mlx_file(
+                        file, mlx_model_id, language, timestamps, verbose,
+                        vad_filter=vad_filter, word_timestamps=word_timestamps,
+                        translate=translate
+                    )
+                    os.makedirs('{}/transcriptions'.format(path), exist_ok=True)
+                    segment_list = []
+                    txt_path = "{}/transcriptions/{}.txt".format(path, title)
+                    srt_path = "{}/transcriptions/{}.srt".format(path, title)
+                    with open(txt_path, 'w', encoding='utf-8') as f, \
+                         open(srt_path, 'w', encoding='utf-8') as srt_f:
+                        f.write(title)
+                        f.write('\n' + '─' * 40 + '\n')
+                        for idx, seg in enumerate(segments, start=1):
+                            if stop_event and stop_event.is_set():
+                                break
+                            text = seg["text"].strip()
+                            if timestamps:
+                                start_ts = str(datetime.timedelta(seconds=seg["start"]))
+                                end_ts   = str(datetime.timedelta(seconds=seg["end"]))
+                                f.write('\n[{} --> {}] {}'.format(start_ts, end_ts, text))
+                            else:
+                                f.write('\n{}'.format(text))
+                            srt_f.write(f'{idx}\n{_srt_timestamp(seg["start"])} --> {_srt_timestamp(seg["end"])}\n{text}\n\n')
+                            f.flush()
+                            srt_f.flush()
+                            if verbose:
+                                print("   [%.2fs → %.2fs] %s" % (seg["start"], seg["end"], seg["text"]))
+                            else:
+                                print("   Transcribed up to %.0fs..." % seg["end"], end='\r')
+                            segment_list.append(seg)
+                    elapsed = time.time() - t_start
+                    elapsed_min = elapsed / 60.0
+                    audio_min   = audio_duration / 60.0
+                    ratio = audio_duration / elapsed if elapsed > 0 else float('inf')
+                    print(f"✅ Done — saved to transcriptions/{title}.txt")
+                    print(f"⏱  Transcribed {audio_min:.1f} min of audio in {elapsed_min:.1f} min  ({ratio:.1f}x realtime)")
+                    files_transcripted.append(segment_list)
+                except Exception as exc:
+                    print(f"⚠  Could not decode '{os.path.basename(file)}', skipping.")
+                    print(f"   Reason: {exc}")
+
+            print(f"\n{SEP}")
+            if files_transcripted:
+                output_text = f"✅ Finished! {len(files_transcripted)} file(s) transcribed.\n   Saved in: {path}/transcriptions"
+            else:
+                output_text = '⚠  No files eligible for transcription — try another folder.'
+            print(output_text)
+            print(SEP)
+            return output_text
+
+    # ── Step 2: Load model (faster-whisper / CTranslate2) ───────────
     print(f"⏳ Loading model '{model}' — downloading if needed...")
-    whisper_model = WhisperModel(model, device=device, compute_type=compute_type)
+    try:
+        whisper_model = WhisperModel(model, device=device, compute_type=compute_type)
+    except Exception as exc:
+        err = str(exc).lower()
+        cuda_runtime_missing = (
+            device == "cuda"
+            and (
+                "libcublas" in err
+                or "libcudnn" in err
+                or "cuda" in err
+                or "cannot be loaded" in err
+                or "not found" in err
+            )
+        )
+        if not cuda_runtime_missing:
+            raise
+        print("⚠  CUDA runtime not available; falling back to CPU (int8).")
+        print(f"   Reason: {exc}")
+        device, compute_type = "cpu", "int8"
+        whisper_model = WhisperModel(model, device=device, compute_type=compute_type)
     print("✅ Model ready!")
     print(SEP)
 
     # ── Step 3: Transcribe files ─────────────────────────────────────
     total_files = len(glob_file)
-    print(f"📂 Found {total_files} item(s) in folder")
+    print(f"📂 Found {total_files} supported media file(s) in folder")
     print(SEP)
 
+    if total_files == 0:
+        output_text = '⚠  No supported media files found — try another folder.'
+        print(output_text)
+        print(SEP)
+        return output_text
+
     files_transcripted = []
     file_num = 0
     for file in glob_file:
+        if stop_event and stop_event.is_set():
+            print("⛔  Transcription stopped by user.")
+            break
         title = os.path.basename(file).split('.')[0]
         file_num += 1
         print(f"\n{'─' * 46}")
         print(f"📄 File {file_num}/{total_files}: {title}")
         try:
+            t_start = time.time()
             segments, info = whisper_model.transcribe(
                 file,
                 language=language,
-                beam_size=5
+                beam_size=5,
+                task="translate" if translate else "transcribe",
+                vad_filter=vad_filter,
+                word_timestamps=word_timestamps,
             )
+            audio_duration = info.duration  # seconds
             # Make folder if missing
             os.makedirs('{}/transcriptions'.format(path), exist_ok=True)
             # Stream segments as they are decoded
             segment_list = []
-            with open("{}/transcriptions/{}.txt".format(path, title), 'w', encoding='utf-8') as f:
+            txt_path = "{}/transcriptions/{}.txt".format(path, title)
+            srt_path = "{}/transcriptions/{}.srt".format(path, title)
+            with open(txt_path, 'w', encoding='utf-8') as f, \
+                 open(srt_path, 'w', encoding='utf-8') as srt_f:
                 f.write(title)
-                for seg in segments:
+                f.write('\n' + '─' * 40 + '\n')
-                    start_ts = str(datetime.timedelta(seconds=seg.start))
+                for idx, seg in enumerate(segments, start=1):
-                    end_ts = str(datetime.timedelta(seconds=seg.end))
+                    if stop_event and stop_event.is_set():
-                    f.write('\n[{} --> {}]:{}'.format(start_ts, end_ts, seg.text))
+                        break
+                    text = seg.text.strip()
+                    if timestamps:
+                        start_ts = str(datetime.timedelta(seconds=seg.start))
+                        end_ts = str(datetime.timedelta(seconds=seg.end))
+                        f.write('\n[{} --> {}] {}'.format(start_ts, end_ts, text))
+                    else:
+                        f.write('\n{}'.format(text))
+                    # Use word-level timestamps for SRT if available
+                    if word_timestamps and hasattr(seg, 'words') and seg.words:
+                        for w_idx, word in enumerate(seg.words, start=1):
+                            w_text = word.word.strip()
+                            if not w_text:
+                                continue
+                            w_start = _srt_timestamp(word.start)
+                            w_end = _srt_timestamp(word.end)
+                            srt_f.write(f'{idx}.{w_idx}\n{w_start} --> {w_end}\n{w_text}\n\n')
+                    else:
+                        srt_f.write(f'{idx}\n{_srt_timestamp(seg.start)} --> {_srt_timestamp(seg.end)}\n{text}\n\n')
                     f.flush()
+                    srt_f.flush()
                     if verbose:
                         print("   [%.2fs → %.2fs] %s" % (seg.start, seg.end, seg.text))
                     else:
                         print("   Transcribed up to %.0fs..." % seg.end, end='\r')
                     segment_list.append(seg)
+            elapsed = time.time() - t_start
+            elapsed_min = elapsed / 60.0
+            audio_min = audio_duration / 60.0
+            ratio = audio_duration / elapsed if elapsed > 0 else float('inf')
             print(f"✅ Done — saved to transcriptions/{title}.txt")
+            print(f"⏱  Transcribed {audio_min:.1f} min of audio in {elapsed_min:.1f} min  ({ratio:.1f}x realtime)")
             files_transcripted.append(segment_list)
-        except Exception:
+        except Exception as exc:
-            print('⚠  Not a valid audio/video file, skipping.')
+            print(f"⚠  Could not decode '{os.path.basename(file)}', skipping.")
+            print(f"   Reason: {exc}")
 
     # ── Summary ──────────────────────────────────────────────────────
     print(f"\n{SEP}")
@@ -147,3 +443,45 @@ def transcribe(path, glob_file, model=None, language=None, verbose=False):
     print(output_text)
     print(SEP)
     return output_text
+
+
+def _transcribe_worker_process(conn, path, glob_file, model, language, verbose, timestamps, vad_filter=False, word_timestamps=False, translate=False):
+    """Child-process entry point for the UI's multiprocessing backend.
+
+    Redirects stdout/stderr → pipe connection so the main process can display
+    output in the console panel.  The main process sends SIGTERM/SIGKILL to
+    stop this process immediately, including any in-progress download or inference.
+    """
+    import sys
+
+    class _PipeWriter:
+        def __init__(self, c):
+            self.c = c
+
+        def write(self, text):
+            if text:
+                try:
+                    self.c.send(text)
+                except Exception:
+                    pass
+
+        def flush(self):
+            pass
+
+    writer = _PipeWriter(conn)
+    sys.stdout = writer
+    sys.stderr = writer
+
+    result = '⚠  No output produced.'
+    try:
+        result = transcribe(path, glob_file, model, language, verbose, timestamps,
+                           vad_filter=vad_filter, word_timestamps=word_timestamps,
+                           translate=translate)
+    except Exception as exc:
+        result = f'⚠  Unexpected error: {exc}'
+    finally:
+        try:
+            conn.send(('__done__', result))
+        except Exception:
+            pass
+        conn.close()