Real-Time Transcription Scripts

This repository contains two Python scripts for real-time audio transcription.

1. transcript_deepgram.py: Uses the Deepgram cloud API for fast and accurate transcription.
2. transcript_local.py: Runs a local Whisper model, optimized for energy efficiency.

Why using deepgram API?

200 usd free usage, each hour is just very low (~0.26 per hour).

Why using local model?

Do not want to configure the API at all

---

1. Deepgram Cloud Transcription (transcript_deepgram.py)

This script streams microphone audio to Deepgram's API for near real-time transcription.

Setup:

1. Install dependencies:

   pip install deepgram-sdk sounddevice numpy requests

2. Add API Key: Create a file named api_config.json and add your key:

   {
     "deepgram_api_key": "YOUR_DEEPGRAM_API_KEY_HERE"
   }

Usage:

Run the script to start transcribing. Press Ctrl+C to stop.

python transcript_deepgram.py

A transcript file will be saved automatically.

---

2. Local Whisper Transcription (transcript_local.py)

This script uses faster-whisper to run transcription locally on your machine, with a focus on minimizing CPU and power usage.

Setup:

1. Install dependencies:

   pip install faster-whisper sounddevice numpy psutil

Usage:

Run the script to start. The base model is used by default. Press Ctrl+C to stop.

python transcript_local.py

A transcript file will be saved automatically.

---

Customization Examples

You can modify the behavior of both scripts with command-line arguments.

Deepgram:

# Use a 3-second chunk size and more sensitive voice detection
python transcript_deepgram.py --chunk-duration 3 --vad-threshold 0.006

Local Whisper:

# Use the 'tiny' model for lowest power usage
python transcript_local.py --model tiny

# Use a larger chunk size and enable low-power sleep mode
python transcript_local.py --chunk-duration 8 --low-power