Voice acting was traditionally one of the biggest single-line costs in an indie game's budget, if you could afford it at all. AI voice generation has gotten good enough that skipping voice acting entirely, for cost reasons, is no longer the only option for a small team. It's not a universal replacement, and knowing the actual boundary saves you from a bad surprise late in production.
Where it's genuinely good enough today
Narration, tutorial voiceover, and dialogue delivered in a calm, consistent register, a system announcer, a codex entry read aloud, sound convincingly natural with current AI voice models. This is also the content type that's most expensive to record traditionally relative to how little players scrutinize it, which makes it the highest-value place to use AI voice first.
Where it still sounds like AI
Combat barks, mid-sentence emotional shifts, a character laughing then suddenly panicking in the same line, and anything requiring precise comedic timing are where current models still noticeably struggle. The tell isn't the voice quality itself anymore, it's the performance: real actors adjust pace and emphasis based on context a model doesn't have, and the gap shows up exactly in the lines that matter most emotionally.
The licensing question is separate from the quality question
Even where the output sounds great, check whether the voice model's license allows the specific commercial use you need, some voice models are licensed per-character or restrict cloning a specific real actor's voice without consent, and that consent question has already produced real lawsuits in the voice-acting industry. Good audio quality and a clean license are two different checks, and you need both before shipping.
A workable hybrid for a small team
Use AI voice for the high-volume, low-emotional-stakes content, narration, tutorials, ambient barks, and budget for a small amount of real human voice talent on the handful of lines that carry your game's emotional weight, a death scene, a key story beat, the villain's final line. That split gets you full voice coverage on a fraction of a traditional budget, without the flat, uncanny delivery landing on the moments players will actually remember.