# Whisper Voice Memo Transcription for Obsidian

## Overview

A simple, free, local transcription setup that:
- Uses OpenAI Whisper (large-v3 model) for high-quality transcription
- Handles French Canadian accent and English seamlessly
- Auto-detects language switches mid-sentence
- Outputs formatted markdown notes to Obsidian
- Runs via conda environment `test_env`

---

## Configuration

| Setting | Value |
|---------|-------|
| Conda Environment | `test_env` |
| Output Directory | `C:\Users\antoi\antoine\My Libraries\Antoine Brain Extension\+\Transcripts` |
| Model | `openai/whisper-large-v3` |
| Supported Formats | mp3, m4a, wav, ogg, flac, webm |

---

## Installation

### Step 1: Activate conda environment

```bash
conda activate test_env
```

### Step 2: Install insanely-fast-whisper

```bash
pip install insanely-fast-whisper
```

### Step 3: Verify installation

```bash
insanely-fast-whisper --help
```

---

## Batch Script: Transcribe.bat

Save this file to your Desktop or a convenient location.

**File:** `Transcribe.bat`

```batch
@echo off
setlocal enabledelayedexpansion

:: ============================================
:: CONFIGURATION - Edit these paths as needed
:: ============================================
set "OUTPUT_DIR=C:\Users\antoi\antoine\My Libraries\Antoine Brain Extension\+\Transcripts"
set "CONDA_ENV=test_env"
set "CONDA_PATH=C:\Users\antoi\anaconda3\Scripts\activate.bat"

:: ============================================
:: MAIN SCRIPT - No edits needed below
:: ============================================

:: Check if file was dragged onto script
if "%~1"=="" (
    echo.
    echo  ========================================
    echo   Voice Memo Transcriber
    echo  ========================================
    echo.
    echo   Drag an audio file onto this script!
    echo   Or paste the full path below:
    echo.
    set /p "AUDIO_FILE=File path: "
) else (
    set "AUDIO_FILE=%~1"
)

:: Generate timestamp for filename
for /f "tokens=1-5 delims=/:.- " %%a in ("%date% %time%") do (
    set "TIMESTAMP=%%c-%%a-%%b %%d-%%e"
)

set "NOTE_NAME=Voice Note %TIMESTAMP%.md"
set "TEMP_FILE=%TEMP%\whisper_output.txt"

echo.
echo  ========================================
echo  Transcribing: %AUDIO_FILE%
echo  Output: %NOTE_NAME%
echo  ========================================
echo.
echo  This may take a few minutes for long recordings...
echo.

:: Activate conda environment and run whisper
call %CONDA_PATH% %CONDA_ENV%
insanely-fast-whisper --file-name "%AUDIO_FILE%" --transcript-path "%TEMP_FILE%" --model-name openai/whisper-large-v3

:: Check if transcription succeeded
if not exist "%TEMP_FILE%" (
    echo.
    echo  ERROR: Transcription failed!
    echo  Check that the audio file exists and is valid.
    echo.
    pause
    exit /b 1
)

:: Create markdown note with YAML frontmatter
echo --- > "%OUTPUT_DIR%\%NOTE_NAME%"
echo created: %date% %time:~0,5% >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo type: voice-note >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo status: raw >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo tags: >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo   - transcript >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo   - voice-memo >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo --- >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo # Voice Note - %date% at %time:~0,5% >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo ## Metadata >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo - **Source file:** `%~nx1` >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo - **Transcribed:** %date% %time:~0,5% >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo --- >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo ## Raw Transcript >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
type "%TEMP_FILE%" >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo --- >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo ## Notes distillees >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo ^<!-- Coller le transcript dans Claude pour organiser et distiller --^> >> "%OUTPUT_DIR%\%NOTE_NAME%"
echo. >> "%OUTPUT_DIR%\%NOTE_NAME%"

:: Cleanup temp file
del "%TEMP_FILE%" 2>nul

echo.
echo  ========================================
echo   DONE!
echo   Created: %NOTE_NAME%
echo   Location: %OUTPUT_DIR%
echo  ========================================
echo.
pause
```

---

## Usage

### Method 1: Drag and Drop
1. Record your voice memo (any app)
2. Drag the audio file onto `Transcribe.bat`
3. Wait for transcription (few minutes for 30min audio)
4. Find your note in Obsidian

### Method 2: Double-click and Paste Path
1. Double-click `Transcribe.bat`
2. Paste the full path to your audio file
3. Press Enter
4. Wait for transcription

---

## Output Format

Each transcription creates a markdown file like this:

```markdown
---
created: 2026-01-15 14:30
type: voice-note
status: raw
tags:
  - transcript
  - voice-memo
---

# Voice Note - 2026-01-15 at 14:30

## Metadata

- **Source file:** `recording.m4a`
- **Transcribed:** 2026-01-15 14:30

---

## Raw Transcript

[Your transcribed text appears here...]

---

## Notes distillees

<!-- Coller le transcript dans Claude pour organiser et distiller -->

```

---

## Processing with Claude

After transcription, use this prompt template to organize your notes:

```
Voici un transcript de notes vocales en français/anglais. 
Peux-tu:

1. Corriger les erreurs de transcription évidentes
2. Organiser par thèmes/sujets
3. Extraire les points clés et action items
4. Reformatter en notes structurées

Garde le contenu original mais rends-le plus lisible.

---

[COLLER LE TRANSCRIPT ICI]
```

---

## Troubleshooting

### "conda is not recognized"
- Verify conda path: `where conda`
- Update `CONDA_PATH` in the script to match your installation

### Transcription takes too long
- The `large-v3` model is accurate but slow on CPU
- For faster (less accurate) results, change model to:
  ```
  --model-name openai/whisper-medium
  ```
  or
  ```
  --model-name openai/whisper-small
  ```

### GPU acceleration
If you have an NVIDIA GPU, install CUDA support:
```bash
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
```

### Wrong language detected
Add language hint to the whisper command:
```bash
insanely-fast-whisper --file-name "audio.mp3" --transcript-path "output.txt" --model-name openai/whisper-large-v3 --language fr
```

---

## Alternative: Python Script Version

For more control or integration with other tools:

**File:** `transcribe.py`

```python
import subprocess
import sys
from datetime import datetime
from pathlib import Path

# Configuration
OUTPUT_DIR = Path(r"C:\Users\antoi\antoine\My Libraries\Antoine Brain Extension\+\Transcripts")
MODEL = "openai/whisper-large-v3"

def transcribe(audio_path: str):
    audio_file = Path(audio_path)
    timestamp = datetime.now().strftime("%Y-%m-%d %H-%M")
    note_name = f"Voice Note {timestamp}.md"
    temp_file = Path.home() / "AppData/Local/Temp/whisper_output.txt"
    
    print(f"\n🎙️ Transcribing: {audio_file.name}")
    print(f"📝 Output: {note_name}\n")
    
    # Run whisper
    subprocess.run([
        "insanely-fast-whisper",
        "--file-name", str(audio_file),
        "--transcript-path", str(temp_file),
        "--model-name", MODEL
    ])
    
    # Read transcript
    transcript = temp_file.read_text(encoding="utf-8")
    
    # Create markdown note
    note_content = f"""---
created: {datetime.now().strftime("%Y-%m-%d %H:%M")}
type: voice-note
status: raw
tags:
  - transcript
  - voice-memo
---

# Voice Note - {datetime.now().strftime("%Y-%m-%d")} at {datetime.now().strftime("%H:%M")}

## Metadata

- **Source file:** `{audio_file.name}`
- **Transcribed:** {datetime.now().strftime("%Y-%m-%d %H:%M")}

---

## Raw Transcript

{transcript}

---

## Notes distillees

<!-- Coller le transcript dans Claude pour organiser et distiller -->

"""
    
    output_path = OUTPUT_DIR / note_name
    output_path.write_text(note_content, encoding="utf-8")
    
    print(f"\n✅ Done! Created: {note_name}")
    print(f"📁 Location: {OUTPUT_DIR}")

if __name__ == "__main__":
    if len(sys.argv) > 1:
        transcribe(sys.argv[1])
    else:
        audio = input("Enter audio file path: ").strip('"')
        transcribe(audio)
```

Run with:
```bash
conda activate test_env
python transcribe.py "path/to/audio.mp3"
```

---

## Next Steps

- [ ] Install `insanely-fast-whisper` in `test_env`
- [ ] Save `Transcribe.bat` to Desktop
- [ ] Test with a short audio clip
- [ ] Pin to taskbar for quick access
- [ ] Set up Claude prompt template for processing

---

## Resources

- [insanely-fast-whisper GitHub](https://github.com/Vaibhavs10/insanely-fast-whisper)
- [OpenAI Whisper](https://github.com/openai/whisper)
- [Whisper model comparison](https://github.com/openai/whisper#available-models-and-languages)