🎙️ Whisper Batch Transcriber

This Python script automates speech-to-text transcription for multiple audio files using OpenAI Whisper.
It scans an audio/ folder for supported files, transcribes them, and saves all results into a single CSV file.

---

Features

• Automatically detects all audio files in the audio/ folder
• Uses OpenAI’s Whisper model for accurate transcription
• Supports multiple audio formats (.mp3, .wav, .m4a, .flac, .ogg, .webm)
• Saves transcripts neatly in a transcripts.csv file
• Flushes results to disk after every file (prevents data loss)
• Displays progress and elapsed time

---

Requirements

• Python 3.8+
• Dependencies:

  pip install openai-whisper
  
  pip install torch  # required backend for Whisper

(If using GPU, install the CUDA-enabled version of PyTorch)

---

Installing the Open-Source Whisper Module (Manual Method)

If you want to manually install Whisper from source, follow these steps:

Download Whisper from GitHub: 🔗 https://github.com/openai/whisper

Extract the downloaded ZIP and navigate to the folder: whisper-main

Install locally using PowerShell or VS Code terminal:

cd C:\Users\User\Downloads\whisper-main
pip install -e .

Alternatively, install dependencies manually:

pip install -r requirements.txt

OR

pip install openai-whisper

✅ This automatically installs all required dependencies, including:

numba
numpy
torch
tqdm
more-itertools
tiktoken

⚠️ Note: On Linux (x86_64), it may also install triton>=2.0.0 if supported by your system.
You do not need to install these manually unless building from source.

---

Installing FFmpeg (Required for Audio Processing)

Download FFmpeg from: 🔗 https://ffmpeg.org/download.html#build-windows

Extract the files and move the folder to:

C:\ffmpeg

Add FFmpeg to your system environment variables:
Go to: Control Panel → System → Advanced System Settings → Environment Variables
Under User variables, find Path

Add:

C:\ffmpeg\bin

Verify installation:

ffmpeg -version

---

📁 Folder Structure

project/
│
├── audio/ # Folder containing your audio files
│ ├── example1.mp3
│ ├── example2.wav
│
├── transcripts.csv # Generated output (after running the script)
│
└── transcribe.py # The main script

---

Usage

Clone the repository:

git clone https://github.com/monojitbgit/whisper-batch-transcriber.git
cd whisper-batch-transcriber

Place your audio files inside the audio/ folder.

Run the script:

python transcribe.py

---

Model Details

By default, the script loads the medium Whisper model:

model = whisper.load_model("medium")

To use a different model:

model = whisper.load_model("small")

Change the transcription language by adjusting:

transcribe_file(file_path, language="en")

(Replace "en" with your desired ISO language code.)

---

Notes

Whisper runs locally — no API key required
For faster performance, use a GPU (CUDA-compatible)
Large models need more memory — try small or base if you get CUDA out of memory

---

⚖️ Whisper Model Comparison

Option	Speed	Accuracy	Model	⏱️ CPU (Intel-i7/Ryzen-7)	⚙️ GPU (RTX 3060/3090/4090)	Size	Speed (Relative)	Accuracy (Relative)	🌐 Multilingual Support
Use "tiny"	Fast	Low accuracy	tiny	~30 sec	~5–10 sec	75 MB	⭐⭐⭐⭐ (Fastest)	⭐ (Basic)	✅ Yes
Use "base"	Fast	Low accuracy	base	~1 min	~10–20 sec	142 MB	⭐⭐⭐ (Fast)	⭐⭐ (Good)	✅ Yes
Use "small" with CPU optimizations	Faster	Good	small	~2–3 min	~20–30 sec	466 MB	⭐⭐ (Medium)	⭐⭐⭐ (Better)	✅ Yes
Use "medium" after reducing file size	Slow	Better	medium	~5–6 min	~40–60 sec	1.5 GB	⭐ (Slower)	⭐⭐⭐⭐ (Very Good)	✅ Yes
Use Google Colab GPU (medium / large)	Super Fast	Best	large	~10–15 min	~1.5–2 min	3.0 GB	⭐ (Slowest)	⭐⭐⭐⭐⭐ (Best)	✅ Yes

---

💡 Recommendations

• 🧠 For quick tests: Use tiny or base
• ⚖️ For balance (speed + accuracy): Use small
• 🗣️ For high-quality transcription: Use medium
• 🚀 For best performance: Use large on GPU or Google Colab

---

License

This project is licensed under the MIT License.
You’re free to use, modify, and distribute it with attribution.