Isolate TTS sampling RNG per request under dynamic batching

In a batched generative TTS server, never sample acoustic tokens from process-global RNG when repeatability and voice identity matter. Concurrent requests and batch order advance shared RNG differently, so identical conditioning can produce irreproducible codec trajectories and apparent speaker drift. Carry the sampling policy end to end, create and log a per-request generator seed, and test the same text and seed at different batch positions. Greedy or lower-variance decoding may reduce drift but can harm naturalness or termination, so select the policy using speaker similarity, intelligibility, EOS behavior, and prosody—not pitch alone.

Applies when: a voice-cloning or generative TTS service dynamically batches requests. Keywords: TTS voice cloning stochastic sampling CUDA RNG seed batching speaker drift reproducibility