Use bundled default audio prompt so conditionals are always prepared

app.py

@@ -19,6 +19,12 @@ subprocess.run(
 from chatterbox_flash import ChatterboxFlashTTS
 
 MODEL_ID = "ResembleAI/chatterbox-flash"
+# Chatterbox-Flash has no built-in voice: `generate()` raises
+# "Conditioning not prepared" unless conditionals exist. Ship a default
+# reference clip so zero-shot generation works without a user upload.
+DEFAULT_AUDIO_PROMPT = os.path.join(
+    os.path.dirname(os.path.abspath(__file__)), "female_shadowheart4.flac"
+)
 _tts = None
 
 
@@ -59,8 +65,13 @@ def generate_tts(
         "backend": "torch",
     }
 
-    if audio_prompt_path:
-        generate_kwargs["audio_prompt_path"] = audio_prompt_path
+    # Fall back to the bundled reference clip when the user supplies none, so
+    # conditionals are always prepared before generation.
+    prompt_path = audio_prompt_path or DEFAULT_AUDIO_PROMPT
+    generate_kwargs["audio_prompt_path"] = prompt_path
+
+    if tts.conds is None:
+        tts.prepare_conditionals(prompt_path, exaggeration=exaggeration)
 
     wav = tts.generate(text_input[:300], **generate_kwargs)
     return (tts.sr, wav.squeeze(0).cpu().numpy())
@@ -92,7 +103,8 @@ with gr.Blocks() as demo:
             ref_wav = gr.Audio(
                 sources=["upload", "microphone"],
                 type="filepath",
-                label="Reference Audio (for voice cloning)",
+                value=DEFAULT_AUDIO_PROMPT,
+                label="Reference Audio (for voice cloning) — defaults to the bundled voice",
             )
 
             with gr.Accordion("Advanced settings", open=False):