Hi everyone,
I wanted to share a small project I just finished. As a bit of a learning exercise, I built a customer service demo system using Resemble AI, and it uses five distinct cloned agent voices. I'm really impressed by how the voices turned out, and I thought I'd walk through how I put it together. Thank you in advance for any thoughts or suggestions you might have.
I started by cloning voices from five of our support team members. The process was straightforward—I uploaded clean audio samples for each person. The quality of the clones is quite good, especially when the voice needs to express empathy or urgency. I integrated the voices into a Python backend using Resemble's API. The core workflow involves taking a customer query, determining the appropriate response and which agent "persona" should deliver it, and then generating the speech on the fly.
For deployment, I containerized the application with Docker and set up a CI/CD pipeline using GitHub Actions. This automatically builds and pushes the image to a registry whenever I update the main branch. It's been a great way to test how these synthetic voices hold up in a semi-realistic, automated flow. The main challenge was managing the latency for real-time generation, but batching some common responses helped a lot.
If anyone else is experimenting with similar setups, I'd love to hear about your experiences, particularly around optimizing pipeline costs or handling different emotional tones effectively.
still learning