Upload

app.py

@@ -59,50 +59,45 @@ model.eval()
    
 @spaces.GPU
 def enhance(filepath):
-    if filepath is None: return None, None
+    # load & resample to model SR
     wav_np, orig_sr = librosa.load(filepath, sr=None)
-    if orig_sr != sr:
-        wav_np = librosa.resample(wav_np, orig_sr, sr)
+    if orig_sr != SR:
+        wav_np = librosa.resample(wav_np, orig_sr, SR)
+    # to tensor + normalize
     wav = torch.from_numpy(wav_np).float().to(device)
     norm = torch.sqrt(len(wav) / torch.sum(wav**2))
     wav = (wav * norm).unsqueeze(0)
-
+    # STFT → model → ISTFT
     amp, pha, _ = mag_phase_stft(wav, **stft_cfg, compress_factor=model_cfg["compress_factor"])
     amp_g, pha_g = model(amp, pha)
     out = mag_phase_istft(amp_g, pha_g, **stft_cfg, compress_factor=model_cfg["compress_factor"])
     out = (out / norm).squeeze().cpu().numpy()
-    if orig_sr != sr:
-        out = librosa.resample(out, sr, orig_sr)
-
-    # write output to temp file
-    out_path = "enhanced_output.wav"
+    # resample back
+    if orig_sr != SR:
+        out = librosa.resample(out, SR, orig_sr)
+    # write to temp file
+    out_path = "enhanced.wav"
     sf.write(out_path, out, orig_sr)
-
-    # plot spectrum
+    # compute spectrogram plot
     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)
+    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 out_path, fig
 
-# --- Gradio Interface ---
 demo = gr.Interface(
     fn=enhance,
-    inputs=gr.Audio(sources=["upload", "microphone"], type="filepath", label="Input Audio"),
+    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
+    description="Upload or record noisy speech; SEMamba enhances it and shows the 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>"
 )
 
 demo.launch()