Remove demo gloss prefill and harden no-detection flow

README.md

@@ -36,12 +36,9 @@ The demo screen is intentionally step-by-step:
 3 Generate speech -> Qwen3-TTS audio
 ```
 
-When the ASL classifier file is missing, the UI exposes a visible debug gloss override instead of
-pretending the model detected words. For an "I love you" demo clip, use:
-
-```text
-I LOVE YOU
-```
+When the ASL classifier file is missing, the app reports `model_missing` and does not invent
+ASL words. An empty-by-default manual gloss override exists only under advanced debug controls
+for testing downstream LLM/TTS behavior.
 
 ## Local checks
 

app.py

@@ -55,6 +55,8 @@ def run_llm_brick(intent_json_text: str) -> tuple[str, str, dict]:
 
 def run_tts_brick(text: str, language: str, speaker: str, instruction: str) -> str:
     try:
+        if (text or "").strip() == "No ASL words were detected yet.":
+            raise ValueError("Analyze ASL did not detect words. Add a real ASL model or use an explicit debug override.")
         return generate_tts(text, language, speaker, instruction)
     except Exception as exc:
         raise gr.Error(f"Qwen3-TTS generation failed: {type(exc).__name__}: {exc}") from exc
@@ -108,12 +110,13 @@ with gr.Blocks(title="SignSpeak Local") as demo:
                 with gr.Column(scale=6, elem_classes=["panel-shell", "input-panel"]):
                     gr.HTML('<div class="section-kicker">01 Capture</div>')
                     full_video_input = build_video_input("Video or camera capture")
-                    full_gloss_override_input = gr.Textbox(
-                        label="Debug gloss override",
-                        value="I LOVE YOU",
-                        lines=1,
-                        info="Used when the ASL classifier is missing or uncertain. Leave empty to use raw model output only.",
-                    )
+                    with gr.Accordion("Advanced debug controls", open=False):
+                        full_gloss_override_input = gr.Textbox(
+                            label="Manual gloss override",
+                            value="",
+                            lines=1,
+                            info="Optional. Use only to test the downstream LLM/TTS when no ASL model is available.",
+                        )
                     with gr.Row(elem_classes=["control-row"]):
                         full_language_input = gr.Dropdown(
                             label="Language",

signspeak/llm.py

@@ -65,7 +65,7 @@ def normalize_llm_output(parsed: dict[str, Any]) -> dict[str, str]:
 
     forbidden_fragments = ["```", '"subtitle"', '"voice_instruction"', "{", "}"]
     if any(fragment in subtitle for fragment in forbidden_fragments):
-        subtitle = "I am happy to see you."
+        subtitle = "I want to say something."
 
     return {
         "subtitle": subtitle,

signspeak/pipeline.py

@@ -67,13 +67,16 @@ def parse_gloss_override(gloss_override: str | None) -> list[str]:
 
 def resolve_video_path(video_file: Any | None) -> Path:
     if video_file:
+        if isinstance(video_file, (list, tuple)):
+            video_file = video_file[0] if video_file else None
         if isinstance(video_file, dict):
             video_file = video_file.get("path") or video_file.get("name") or video_file.get("video")
         elif hasattr(video_file, "path"):
             video_file = video_file.path
         elif hasattr(video_file, "name"):
             video_file = video_file.name
-        return Path(video_file)
+        if video_file:
+            return Path(video_file)
     if DEFAULT_VIDEO_PATH.exists():
         return DEFAULT_VIDEO_PATH
     return create_synthetic_demo_video()

tests/test_asl_pipeline.py

@@ -46,6 +46,15 @@ def test_resolve_video_path_accepts_gradio_dict_payload(tmp_path):
     assert resolved == video_path
 
 
+def test_resolve_video_path_accepts_gradio_tuple_payload(tmp_path):
+    video_path = tmp_path / "capture.mp4"
+    video_path.write_bytes(b"demo")
+
+    resolved = resolve_video_path((str(video_path),))
+
+    assert resolved == video_path
+
+
 def test_parse_gloss_override_normalizes_words():
     assert parse_gloss_override("i love,you") == ["I", "LOVE", "YOU"]
 

tests/test_llm_parsing.py

@@ -37,7 +37,7 @@ def test_normalize_llm_output_blocks_json_subtitle():
         }
     )
 
-    assert normalized["subtitle"] == "I am happy to see you."
+    assert normalized["subtitle"] == "I want to say something."
     assert normalized["voice_instruction"] == "Speak warmly."