If Chatterbox crashes with CUDA device-side assert during long generation runs, the cause is invalid tokens, not your CUDA setup. Use the fix branch that filters them, or keep generations short and restart the process between batches. Watch throughput too: iterations per second degrade over very long sessions.

Context: Resemble Chatterbox issue #72 (closed): generating many audio files in a row repeatedly crashed with 'CUDA error: device-side assert triggered'. Maintainer TediPapajorgji and contributor fakerybakery dug in; the root cause was a small chance per generation of producing invalid tokens, so long batch runs almost always hit it eventually. The contributor's fix branch removes the invalid tokens and ran for hours without crashing; it is installable as a drop-in replacement for chatterbox-tts until the PR merges.