If you’ve ever done long-form audio, you know this pain.
You spend ages generating a ten-minute voiceover, listen through the whole thing, and then — one word cracks, or one word gets mispronounced. Now what? You re-generate the entire thing. Your credits just burned, like pulling a gacha card and watching your heart break.
This is exactly where I keep getting stuck when I’m working on voice for our AI team. ElevenLabs is great, no doubt, but that subscription-plus-credits model — honestly, it wears on you after a while. You’re not paying for “content.” You’re paying for “trial and error.”
And today, there’s a crack in that wall.
Fish Audio just dropped a voice model called Drama 3, and it’s open-weight (more on why that’s different from “open source” in a second). The wildest part: on Artificial Analysis’s benchmark, it beats ElevenLabs straight-up on pronunciation accuracy. The launch video even features Uncle Roger “directing” an AI clone of his own voice (with his consent) — that alone makes for a good show.
But what actually made me sit up was the word-level editing.
You know that nightmare where one wrong word means re-recording the whole thing? Drama 3’s approach: you mark the word you want to fix, type in the correct one, and the model only regenerates that single word — automatically matching the tone, emotion, and pacing around it. Everything else you’d already approved stays exactly as it was.
This is actually hitting a real pain point for creators — it’s solving the cost of redoing work, not just bolting on another flashy feature. Though I’ll be honest: word-level editing is still invite-only testing right now, not something everyone can use immediately. But the fact that they’re even going in this direction is genuinely interesting.
Emotion control is there too. Like ElevenLabs, you can embed tags in your text telling it to sigh, laugh, whisper, choke up a little — crucial for character work, dialogue, or any longer content that needs emotional range.
Now, what’s the actual difference between “open-weight” and “free and open source”? I want to be clear on this, because it’s easy to mix up, and it’s exactly the kind of thing a flashy headline glosses over.
Drama 3’s weights are on HuggingFace and the code is on GitHub, but it runs under Fish Audio’s own community license: free for research and non-commercial use. For commercial use, it’s only free if you (and your affiliated entities) have annual revenue under $500K USD — above that, you need to negotiate a separate license. It also explicitly bans voice cloning without consent.
So it’s not the “completely free, use it however you want” kind of open source. But for the vast majority of individual creators, small teams, and people just tinkering with agents on their own, that threshold is practically free. That’s the part that matters.
I should also temper expectations a bit. This is a 35B model (MoE architecture, so only a small slice of the parameters activates per run), and the official numbers show it decoding 271 frames per second on a single stream, on a single H200 GPU. In other words, this isn’t the kind of thing that responds in milliseconds on your phone — you need real hardware to run it smoothly. If you see people online claiming it has “millisecond-level response,” I checked and the official docs don’t say that — don’t take it at face value.
None of that changes the bigger picture though: ElevenLabs’ paywall moat, built on closed weights plus subscriptions, just took a real hit today.
For anyone building agents, for creators doing long-form audio, this is good news. You’ve now got another option — one where you don’t have to watch your credits burn every time a word cracks. The moat has a crack in it now, and water’s going to keep seeping in. What happens next depends on whether Fish Audio can keep this up, and whether ElevenLabs is willing to drop that stance that’s been putting so much pressure on everyone.
I’m going to go play around with it myself and report back once I’ve got something real to say.