Hey everyone, been lurking for a bit but finally have something to post about. I'm working on a side project where we need dynamic voice responses, and I've been testing a bunch of TTS APIs. Wanted to share my experience with Resemble AI.
For creating pre-recorded voice clips or short dialogues, Resemble is actually pretty solid. The voice cloning quality from a small sample is impressive, and the interface for building those clips is intuitive. I got a demo up and running for static content in an afternoon.
But here's where I hit a wall. We needed real-time synthesis via their API for personalized responses. The latency is all over the place. Sometimes it's okay, but other times there's a noticeable lag that just breaks the flow of a conversation. I also ran into weird glitches with longer generation tasks—like the stream would just hang. My orchestration layer (I'm learning Airflow) kept flagging these tasks as failures due to timeouts, which was a pain to handle.
As someone still getting their head around production data pipelines, this feels like a big red flag. If the API isn't consistently reliable, how do you build a robust service around it? Do you just wrap every call in a ton of retry logic and hope for the best? 😅
Has anyone else tried using their real-time API in a live application? Am I missing something in my setup, or is this a known thing? I really wanted to like it because the voice quality is good, but the inconsistency is tough.
-- rookie
rookie