density-speaker-label

Customer Service Call Transcription and Speaker Diarization

This project provides an program to transcribe customer service call recordings and identify speakers (diarization), producing a structured, readable conversation text file.

Runtime Technical Requirements

• Programming Language: TypeScript
• Runtime Environment: Node.js (Dockerized for consistency)
• Host Machine: MacOS (but Docker ensures cross-platform compatibility)

Note:
Docker is used to eliminate dependency issues across different Node.js versions. Running inside containers guarantees consistent behavior on any machin or chip.

---

Transcription Model: Whisper

The transcription is powered by Whisper, OpenAI’s state-of-the-art speech-to-text model for.

• Why Whisper?
  I've consistently found Whisper to deliver decent accuracy over time, However you can try other models if you so desire it.

• Requirements:
  An OpenAI account and API key are needed. You can obtain one through the OpenAI Developer Portal.

• Additional Notes:

  • A 100% transcription success rate was achieved in testing.
  • Whisper offers multiple models; for larger or self-hosted setups, see the official GitHub repository.

---

🗣 Speaker Diarization: Pyannote

Since Whisper does not currently support diarization, Pyannote is used to assign transcribed text to individual speakers.

• About Pyannote:
  An open-source Python toolkit for speaker diarization, built on PyTorch. It provides highly accurate diarization services both through self-hosting and SaaS offerings.

• API Access:
  To use the hosted Pyannote service, create an API key following their documentation and pricing page.

• Self-Hosting Option:
  If needed, Pyannote models can be self-hosted by following the instructions on their GitHub repository.

---

Setup Instructions

1. Create a .env File

   In the root of the project folder, create a .env file and add your API keys:

   OPENAI_API_KEY=your_openai_api_key
   PYANNOTE_API_KEY=your_pyannote_api_key
   

2. Run the Application via Docker Compose

   In the project directory, run:

   docker-compose up --build

   Docker will build and start the container automatically.

3. Modify the Audio File (Optional)

   • If you wish to use a different audio file:
     • Open src/index.ts
     • Update the audioUrl variable with your own CDN-hosted audio link.

   Important:
   Ensure the audio URL is a publicly accessible CDN link (such as Cloudinary or AWS S3) and not a private storage link (like Dropbox or Google Drive).

---

Output

• A plain text file containing the full conversation is generated after processing.
• Speakers are clearly identified (e.g., Agent, Caller) and mapped to each transcription segment.

---

Contact

For any questions or collaboration requests, feel free to reach out:

Email: obakunleoluseye@gmail.com