diff --git a/stacks/chatterbox-fast/app.py b/stacks/chatterbox-fast/app.py
index 758456b..7b4e476 100644
--- a/stacks/chatterbox-fast/app.py
+++ b/stacks/chatterbox-fast/app.py
@@ -206,6 +206,7 @@ class TTSRequest(BaseModel):
     top_p: float = 0.95
     top_k: int = 1000
     repetition_penalty: float = 1.2
+    seed: int = 0  # 0 ⇒ random; a fixed seed repeats a one-shot take (see note below)
     # Scheduler overrides (None ⇒ ChunkConfig defaults).
     margin: float | None = Field(default=None)
     margin_first: float | None = Field(default=None)
@@ -279,6 +280,13 @@ def tts(req: TTSRequest) -> StreamingResponse:
         # workload); concurrent callers queue rather than corrupt conditionals.
         with engine.lock:
             engine._prepare(voice_path, exaggeration=req.exaggeration)
+            # Seed once per request (under the lock). One-shot is then reproducible
+            # for a fixed seed + params; streaming is NOT — adaptive-chunk boundaries
+            # depend on live-measured RTF (wall-clock), so chunk splits vary run to run.
+            if req.seed:
+                torch.manual_seed(req.seed)
+                if DEVICE.startswith("cuda"):
+                    torch.cuda.manual_seed_all(req.seed)
             t_req = time.perf_counter()
             first_audio_ms: float | None = None