Gradio 6 theme/css on launch, tighter GPU duration, cleaner edit transcript

app.py

@@ -8,6 +8,7 @@ Three capabilities of https://huggingface.co/FireRedTeam/FireRedTTS3 :
 
 import functools
 import os
+import re
 import urllib.request
 
 import spaces  # noqa: F401  (must precede torch / CUDA imports)
@@ -134,6 +135,16 @@ def _to_gradio_audio(audio: torch.Tensor, sr: int):
     return sr, (x * 32767.0).astype(np.int16)
 
 
+_EDIT_MASK_RE = re.compile(r"<\|edit\|>(?:<\|frame_patch\|>)*<\|end_edit\|>")
+
+
+def _pretty_edit_text(text: str) -> str:
+    """The model marks the re-synthesized span with edit/frame-patch tokens."""
+    text = _EDIT_MASK_RE.sub(" ⟨edited span⟩ ", text or "")
+    text = re.sub(r"<\|[^|]*\|>", "", text)
+    return re.sub(r"\s+", " ", text).strip()
+
+
 def _check_text(text: str, what: str = "Text"):
     text = (text or "").strip()
     if not text:
@@ -147,7 +158,7 @@ def _check_text(text: str, what: str = "Text"):
 # --------------------------------------------------------------------------- #
 # Inference
 # --------------------------------------------------------------------------- #
-@spaces.GPU(duration=150)
+@spaces.GPU(duration=60)
 def voice_clone(
     prompt_audio,
     prompt_text,
@@ -179,7 +190,7 @@ def voice_clone(
     return _to_gradio_audio(gen_audio, gen_sr)
 
 
-@spaces.GPU(duration=150)
+@spaces.GPU(duration=60)
 def voice_design(
     instruction,
     text,
@@ -203,7 +214,7 @@ def voice_design(
     return _to_gradio_audio(gen_audio, gen_sr), (gen_text or "").strip()
 
 
-@spaces.GPU(duration=150)
+@spaces.GPU(duration=60)
 def semantic_edit(
     audio_in,
     instruction,
@@ -223,7 +234,7 @@ def semantic_edit(
         inference_cfg=float(inference_cfg),
         seed=int(seed),
     )
-    return _to_gradio_audio(gen_audio, gen_sr), (gen_text or "").strip()
+    return _to_gradio_audio(gen_audio, gen_sr), _pretty_edit_text(gen_text)
 
 
 def compose_acoustic_instruction(attribute: str, value: float) -> str:
@@ -236,7 +247,7 @@ def compose_acoustic_instruction(attribute: str, value: float) -> str:
     return f"shift the pitch by {steps} step{'' if abs(steps) == 1 else 's'}"
 
 
-@spaces.GPU(duration=150)
+@spaces.GPU(duration=60)
 def acoustic_edit(
     audio_in,
     attribute="Speed",
@@ -276,7 +287,7 @@ CSS = """
 .dark .gradio-container {color: var(--body-text-color);}
 """
 
-with gr.Blocks(theme=gr.themes.Citrus(), css=CSS, title="FireRedTTS3") as demo:
+with gr.Blocks(title="FireRedTTS3") as demo:
     gr.Markdown(
         """
         # 🔥 FireRedTTS3 — Unified Speech Generation & Editing
@@ -573,4 +584,4 @@ with gr.Blocks(theme=gr.themes.Citrus(), css=CSS, title="FireRedTTS3") as demo:
     )
 
 if __name__ == "__main__":
-    demo.queue().launch(mcp_server=True)
+    demo.queue().launch(theme=gr.themes.Citrus(), css=CSS, mcp_server=True)