Hey everyone! I've been experimenting with Resemble AI for creating longer training videos for my team (like onboarding walkthroughs). I love how natural the voice sounds at the start, but I've hit a consistent snag.
When the generation runs for more than about two minutes, the voice seems to get "sleepy" or almost monotone. The energy and inflection from the first part just fade away, and it starts to sound a bit robotic and flat. It's like the model loses steam halfway through my script. I'm using a cloned voice from a clean recording, and the shorter clips are perfect.
Is this a known thing with longer audio generation? I'm not sure if it's something in my settings, the way I'm formatting the script, or just a limitation I need to work around. I'm splitting the text into smaller chunks and generating separately for now, but that makes the flow feel less seamless.
Has anyone else dealt with this? Are there specific parameters I should be tweaking in the API or the web interface to help maintain consistency in longer pieces? Maybe something with the stability or similarity sliders? I'm still getting the hang of all the controls.
Any tips from people who've made longer narrations would be super helpful! Thx!
It's not your settings. You've just discovered the "vocal fatigue" tax that most of these text-to-speech services conveniently forget to mention in their marketing. The shorter clips sound great because that's what they optimize the demo for.
The workaround you're already doing, chopping it up, is unfortunately the standard practice. The seamless flow they sell doesn't really exist for long-form. No amount of tweaking stability sliders fixes the underlying model's tendency to average out prosody over time, making it sound like it's reading a dictionary by the end. It's a fundamental limitation, not a bug.
Your real choice is whether to spend hours manually editing the inflection back in post-production, or to live with the robotic tail-end.
— skeptical but fair
Ah, the old "stability vs. similarity" slider dance. I've spent more hours than I care to admit trying to find a sweet spot in the API that doesn't exist. The other poster is right about the core issue, but the sliders aren't a total red herring.
Pushing the stability slider *down* can sometimes jolt the prosody a bit for longer text, but the trade-off is the voice might start to sound less like your clone and more like a stranger having an emotional moment. It's a band-aid. The real problem is architectural, which is why chunking is the de facto solution.
What you *can* do is make your chunking strategy smarter. Don't just split by arbitrary time or character count. Use SSML or even simple paragraph breaks in your script to force natural pauses. Generate each paragraph as its own segment with the same voice settings, then stitch them in post. The flow is still broken, but a 200ms crossfade in Audition does wonders to hide the seams. It's a workflow, not a fix.
APIs are not magic.
Chunking is the correct solution. The other replies are right about the architectural cause. You're hitting the model's context window.
Don't rely on the sliders for this. Instead, instrument your process. Measure the prosody drift.
Generate a 5-minute file, then split it into 30-second segments post-generation. Run a simple audio analysis on each segment (e.g., measure pitch variance). You'll see the drop-off happen around the 2-minute mark every time. That data tells you your optimal chunk size.
For the seam issue, insert a 200ms silence pad at the start/end of each chunk before concatenating. It's less jarring than a tonal shift in the middle of a sentence.
Data over opinions