deepgram-stt

Speech to text using Deepgram API with speaker diarization.

Installation

pip install -r requirements.txt

Configuration

Create a .env file with your Deepgram API key:

DEEPGRAM_API_KEY=your_api_key_here

Usage

1. Transcribe audio

python speech_to_text.py <audio_file> [-l <language>]

Examples:

# Transcribe English audio (default)
python speech_to_text.py interview.wav

# Transcribe Russian audio
python speech_to_text.py podcast.mp3 -l ru

Output: Creates output/<filename>/ folder with:

• <filename>.json - Raw API response
• <filename>_speakers.txt - Speaker-labeled transcript
• <filename>_youtube.txt - YouTube timestamp format

2. Clean transcript

Merge consecutive same-speaker lines into cleaner format:

python clean_transcript.py <folder>

Example:

# Clean transcript for interview.wav (uses output/interview/)
python clean_transcript.py interview

Output: Creates output/<folder>/<folder>_clean.txt

Language codes

• en - English (default)
• uk - Ukrainian
• ru - Russian
• de - German
• fr - French
• es - Spanish

Deepgram Real-time STT with Diarization

Real-time speech-to-text transcription with speaker diarization using Deepgram's WebSocket API.

Features

• Real-time microphone transcription
• Speaker diarization (identifies different speakers)
• Color-coded speaker labels in terminal
• Interim results (see words as you speak)
• Automatic transcript saving

Setup

1. Install dependencies:

pip install -r requirements.txt

2. Create .env file with your Deepgram API key:

DEEPGRAM_API_KEY=your_api_key_here

Or copy from the deepgram-stt folder:

cp ../deepgram-stt/.env .

Usage

Basic usage (English):

python realtime_stt.py

With different language:

python realtime_stt.py -l ru    # Russian
python realtime_stt.py -l de    # German
python realtime_stt.py -l fr    # French

Without saving transcript:

python realtime_stt.py --no-save

List available audio devices:

python realtime_stt.py --list-devices

Use specific audio device:

python realtime_stt.py -d 2     # Use device ID 2

Output

Transcripts are saved to the output/ folder:

• YYYYMMDD_HHMMSS_realtime.txt - Speaker-labeled transcript
• YYYYMMDD_HHMMSS_realtime.json - Full metadata with timestamps

Controls

• Ctrl+C - Stop recording and save transcript

Terminal Display

Each speaker is shown with a different color:

• Speaker 0: Cyan
• Speaker 1: Green
• Speaker 2: Yellow
• Speaker 3: Magenta
• etc.

Interim results are shown with ... suffix and update in place. Final results are printed on a new line.