Upload

app.py

@@ -58,13 +58,13 @@ model.eval()
 
    
 @spaces.GPU
-def enhance(audio):
-    if audio is None: return None, None
-    orig_sr, wav_np = audio
+def enhance(filepath):
+    if filepath is None: return None, None
+    wav_np, orig_sr = librosa.load(filepath, sr=None)
     if orig_sr != sr:
         wav_np = librosa.resample(wav_np, orig_sr, sr)
     wav = torch.from_numpy(wav_np).float().to(device)
-    norm = torch.sqrt(len(wav) / torch.sum(wav ** 2))
+    norm = torch.sqrt(len(wav) / torch.sum(wav**2))
     wav = (wav * norm).unsqueeze(0)
 
     amp, pha, _ = mag_phase_stft(wav, **stft_cfg, compress_factor=model_cfg["compress_factor"])
@@ -74,42 +74,35 @@ def enhance(audio):
     if orig_sr != sr:
         out = librosa.resample(out, sr, orig_sr)
 
-    # spectrogram
+    # write output to temp file
+    out_path = "enhanced_output.wav"
+    sf.write(out_path, out, orig_sr)
+
+    # plot spectrum
     D = librosa.stft(out, n_fft=1024, hop_length=512)
     S_db = librosa.amplitude_to_db(np.abs(D), ref=np.max)
     fig, ax = plt.subplots(figsize=(6, 3))
     librosa.display.specshow(S_db, sr=orig_sr, hop_length=512, x_axis='time', y_axis='hz', ax=ax)
     ax.set_title("Enhanced Spectrogram")
     plt.colorbar(format="%+2.0f dB", ax=ax)
-    return (orig_sr, out), fig
-
-# --- Layout with Blocks ---
-with gr.Blocks(css=".gr-box {border: none !important}") as demo:
-    gr.Markdown("<h1 style='text-align: center;'>🎧 <a href='https://github.com/RoyChao19477/SEMamba' target='_blank'>SEMamba</a>: Speech Enhancement</h1>")
-    gr.Markdown("Enhance real-world noisy speech using Mamba. Upload or record an audio clip and view the spectrogram.")
-
-    with gr.Row():
-        with gr.Column():
-            audio_input = gr.Audio(sources=["upload", "microphone"], type="numpy", label="Upload or Record", elem_id="input-audio")
-            run_btn = gr.Button("Enhance Now 🚀", variant="primary")
-
-        with gr.Column():
-            enhanced_audio = gr.Audio(label="Enhanced Output", type="numpy")
-            spec_plot = gr.Plot(label="Spectrogram")
-
-    run_btn.click(enhance, inputs=audio_input, outputs=[enhanced_audio, spec_plot])
-
-    gr.Examples(
-        examples=[
-            ["examples/noisy_sample_16k.wav"],
-        ],
-        inputs=audio_input,
-        outputs=[enhanced_audio, spec_plot],
-        fn=enhance,
-        cache_examples=True,
-        label="📂 Try These Examples"
-    )
-
-    gr.Markdown("<p style='text-align: center'><a href='https://arxiv.org/abs/2405.15144' target='_blank'>📄 SEMamba: Mamba for Long-Context Speech Enhancement (SLT 2024)</a></p>")
+
+    return out_path, fig
+
+# --- Gradio Interface ---
+demo = gr.Interface(
+    fn=enhance,
+    inputs=gr.Audio(sources=["upload", "microphone"], type="filepath", label="Input Audio"),
+    outputs=[
+        gr.Audio(label="Enhanced Audio", type="filepath"),
+        gr.Plot(label="Spectrogram")
+    ],
+    title="<a href='https://github.com/RoyChao19477/SEMamba' target='_blank'>SEMamba</a>: Speech Enhancement",
+    description="Upload or record a noisy audio file and SEMamba will enhance the speech and show its spectrogram.",
+    article="<p style='text-align: center'><a href='https://arxiv.org/abs/2405.15144' target='_blank'>📄 SEMamba: Mamba for Long-Context Speech Enhancement (SLT 2024)</a></p>",
+    examples=[
+        ["examples/noisy_sample_16k.wav"]
+    ],
+    cache_examples=True
+)
 
 demo.launch()