fix: Chatterbox uses separate classes per variant, remove turbo

ChatterboxTTS and ChatterboxMultilingualTTS are separate classes.
Turbo variant doesn't exist in chatterbox-tts 0.1.7.
Multilingual generate() requires language_id parameter.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

kischdle/llmux/llmux/backends/chatterbox_tts.py

@@ -1,4 +1,5 @@
 import asyncio
+import gc
 import io
 import logging
 
@@ -24,25 +25,26 @@ class ChatterboxTTSBackend(BaseBackend):
         logger.info(f"Loading Chatterbox {variant} to {device}")
 
         def _load():
-            from chatterbox.tts import ChatterboxTTS
-            if variant == "turbo":
-                model = ChatterboxTTS.from_pretrained(device=device, variant="turbo")
-            elif variant == "multilingual":
-                model = ChatterboxTTS.from_pretrained(device=device, variant="multilingual")
+            if variant == "multilingual":
+                from chatterbox import ChatterboxMultilingualTTS
+                return ChatterboxMultilingualTTS.from_pretrained(device=device)
             else:
-                model = ChatterboxTTS.from_pretrained(device=device)
-            return model
+                from chatterbox.tts import ChatterboxTTS
+                return ChatterboxTTS.from_pretrained(device=device)
 
         loop = asyncio.get_event_loop()
         model = await loop.run_in_executor(None, _load)
-        self._loaded[model_id] = {"model": model, "device": device}
+        self._loaded[model_id] = {"model": model, "variant": variant, "device": device}
 
     async def unload(self, model_id: str) -> None:
         if model_id not in self._loaded:
             return
         entry = self._loaded.pop(model_id)
         del entry["model"]
+        del entry
+        gc.collect()
         torch.cuda.empty_cache()
+        logger.info(f"Unloaded Chatterbox {model_id}")
 
     async def generate(self, model_id, messages, params, stream=False, tools=None):
         raise NotImplementedError("TTS backend does not support chat generation")
@@ -50,9 +52,15 @@ class ChatterboxTTSBackend(BaseBackend):
     async def synthesize(self, model_id: str, text: str, voice: str = "default") -> bytes:
         entry = self._loaded[model_id]
         model = entry["model"]
+        variant = entry["variant"]
 
         def _synthesize():
-            wav = model.generate(text)
+            if variant == "multilingual":
+                # Default to English; voice param could encode language
+                lang = "en" if voice == "default" else voice
+                wav = model.generate(text, language_id=lang)
+            else:
+                wav = model.generate(text)
             buf = io.BytesIO()
             sf.write(buf, wav.cpu().numpy().squeeze(), samplerate=24000, format="WAV")
             buf.seek(0)