A Python library for transcribing media content with speaker diarization support. This package provides high-quality transcription using WhisperX with automatic speaker detection and labeling.

• High-quality transcription using WhisperX
• Automatic speaker diarization
• Support for multiple media sources
• GPU acceleration support
• Flexible output formats
• Type-safe interfaces

pip install diarized-transcriber

• Python 3.10 or later
• CUDA-capable GPU
• PyTorch with CUDA support
• HuggingFace account for pyannote.audio access

Before running the examples below you must export an authentication token for the diarization pipeline. This token is referenced in diarized_transcriber/transcription/diarization.py (see lines 43–55) and is required for downloading the pretrained pyannote.audio model.

export HF_TOKEN="<your-huggingface-token>"

from diarized_transcriber import TranscriptionEngine, MediaContent, MediaSource

# Initialize the engine
engine = TranscriptionEngine(model_size="base")

# Create media content object
content = MediaContent(
    id="example-1",
    title="Example Media",
    media_url="https://example.com/media.mp3",
    source=MediaSource(type="podcast")
)

# Perform transcription
result = engine.transcribe(content)

# Format the result
from diarized_transcriber.utils.formatting import format_transcript
transcript = format_transcript(
    result,
    output_format="text",
    group_by_speaker=True
)

print(transcript)

Set up your HuggingFace token for pyannote.audio access:

export HF_TOKEN='your-huggingface-token'

Ensure CUDA is properly configured for your system. A good way to check if CUDA is working is to run nvidia-smi in your terminal.

The TranscriptionEngine can be configured with different model sizes:

• tiny: Fastest, lowest accuracy
• base: Good balance of speed and accuracy
• small: Better accuracy, slower than base
• medium: High accuracy, slower
• large: Highest accuracy, slowest

Example configuration:

engine = TranscriptionEngine(
    model_size="medium",
    compute_type="float16"  # or "float32" for higher precision
)

The library supports multiple output formats.

transcript = format_transcript(
    result,
    output_format="text",
    include_timestamps=True,
    timestamp_format="HH:MM:SS.mmm",
    group_by_speaker=True
)

transcript_dict = format_transcript(
    result,
    output_format="dict",
    include_timestamps=True
)

Install the optional dependencies:

pip install diarized-transcriber[server]

Start the service:

python -m diarized_transcriber.api.server

The library provides specific exceptions for different error cases:

• GPUConfigError: GPU-related issues
• ModelLoadError: Model loading failures
• AudioProcessingError: Audio processing problems
• TranscriptionError: General transcription failures
• DiarizationError: Speaker diarization issues

Contributions are welcome! Please see our contributing guidelines for more details.

MIT License - see LICENSE file for details.