ComfyUI Custom Nodes: Speaker Diarization and Whisper Integration

Overview

This repository provides custom nodes for ComfyUI designed to process audio files, performing speaker diarization and integrating speaker data into whisper-transcribed segments. These nodes utilize the PyAnnote library for speaker identification and pandas for efficient data handling.

---

Custom Nodes

1. Speaker Diarization Node

Description:
Performs speaker diarization on an input audio file, identifying and segmenting speech by different speakers.

• Category: Audio Processing
• Node Name: Speaker Diarization

Inputs:

• audio: (AUDIO)
  The audio data to process.
• hf_token: (STRING)
  A Hugging Face authentication token for accessing the PyAnnote model.

Outputs:

• speaker_segments: (List[Dict])
  A list of dictionaries containing speaker segments with start, end, and speaker information.

Usage Example:
After connecting an audio input, this node identifies where each speaker starts and ends in the audio file.

---

2. Whisper Segments to Speaker Node

Description:
Adds speaker labels to segments generated by a Whisper speech-to-text model, aligning them based on time overlaps with diarized segments.

• Category: Audio Processing
• Node Name: Whisper Segments to Speaker

Inputs:

• whisper_segments: (whisper_alignment)
  Transcription segments generated by Whisper, including start and end times.
• speaker_segments: (speaker_segments)
  Output from the Speaker Diarization Node, providing time-aligned speaker information.

Outputs:

• segments_alignment: (whisper_alignment)
  The original Whisper segments enriched with speaker labels.

Usage Example:
Connect the output of the Speaker Diarization Node and Whisper-transcribed segments. This node will align speaker data, allowing for a detailed transcription with speaker differentiation.

---

Dependencies

Ensure the following libraries are installed:

• PyAnnote Audio: For speaker diarization.
• pandas: For data manipulation.
• numpy: For numeric operations.
• torchaudio: For audio processing.
• pydub: For audio format handling.

Install dependencies using:

pip install pyannote.audio pandas numpy torchaudio pydub

---

Integration with ComfyUI

1. Place Node Files:
   Add the provided Python file to the custom nodes directory in your ComfyUI setup.

2. Register the Nodes:
   ComfyUI automatically detects nodes from the NODE_CLASS_MAPPINGS. Ensure the structure includes:

   NODE_CLASS_MAPPINGS = {
       "Speaker Diarization": SpeakerDiarizationNode,
       "Whisper Segments to Speaker": WhisperDiarizationNode
   }

3. Node Interface:
   In the ComfyUI interface, locate the nodes under the Audio Processing category. Connect them as needed to process audio inputs and transcriptions.

---

Configuration

Hugging Face Token

To access the PyAnnote model, you need a Hugging Face token. Sign up or log in to Hugging Face and generate an access token from your account settings.

---

License

This project is open-source. Refer to the LICENSE file for details.

---

Contact

For issues or feature requests, please submit them via the repository's issue tracker.

---

Happy audio processing! 🎧