Upload

app.py

@@ -24,12 +24,6 @@ from models.stfts    import mag_phase_stft, mag_phase_istft
 from models.generator import SEMamba
 from models.pcs400   import cal_pcs
 
-# download model files from your HF repo
-#ckpt  = hf_hub_download("rc19477/Speech_Enhancement_Mamba",
-#                        "ckpts/SEMamba_advanced.pth")
-#cfg_f = hf_hub_download("rc19477/Speech_Enhancement_Mamba",
-#                        "recipes/SEMamba_advanced.yaml")
-
 ckpt = "ckpts/SEMamba_advanced.pth"
 cfg_f = "recipes/SEMamba_advanced.yaml"
 
@@ -60,38 +54,49 @@ load_model()
    
 @spaces.GPU
 def enhance(filepath):
-    wav_np, orig_sr = librosa.load(filepath, sr=None)
+    # load & (if needed) resample to model SR
+    wav, 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))
-    wav = (wav * norm).unsqueeze(0)
-    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()
+        wav = librosa.resample(wav, orig_sr, SR)
+    # normalize → tensor
+    x = torch.from_numpy(wav).float().to(device)
+    norm = torch.sqrt(len(x)/torch.sum(x**2))
+    x = (x*norm).unsqueeze(0)
+    # STFT → model → ISTFT
+    amp,pha,_ = mag_phase_stft(x, **stft_cfg, compress_factor=model_cfg["compress_factor"])
+    amp2,pha2 = model(amp, pha)
+    out = mag_phase_istft(amp2, pha2, **stft_cfg, compress_factor=model_cfg["compress_factor"])
+    out = (out/norm).squeeze().cpu().numpy()
+    # back to original rate
     if orig_sr != SR:
         out = librosa.resample(out, SR, orig_sr)
+    # write file
     sf.write("enhanced.wav", out, orig_sr)
-
+    # build spectrogram
     D = librosa.stft(out, n_fft=1024, hop_length=512)
     S = librosa.amplitude_to_db(np.abs(D), ref=np.max)
-    fig, ax = plt.subplots(figsize=(6, 3))
+    fig, ax = plt.subplots(figsize=(6,3))
     librosa.display.specshow(S, 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 "enhanced.wav", fig
 
-    return out, fig
+ABOUT = """
+# SEMamba: Speech Enhancement
+A Mamba-based model that denoises real-world audio.
+Upload or record a noisy clip and click **Enhance** to hear + see its spectrogram.
+"""
 
-# --- Build demo ---
 with gr.Blocks() as demo:
-    audio_in = gr.Audio(label="Input Audio", type="filepath", interactive=True)
-    run_btn = gr.Button("Enhance", variant="primary")
-    enhanced_audio = gr.Audio(label="Enhanced Audio", type="filepath", interactive=False)
-    spectrogram = gr.Plot(label="Spectrogram")
-
-    run_btn.click(fn=enhance, inputs=audio_in, outputs=[enhanced_audio, spectrogram])
-
-    gr.Markdown("Unofficial demo by [yourname](https://github.com/RoyChao19477)")
-
-demo.launch()
+    gr.Markdown(ABOUT)
+    audio_in     = gr.Audio(sources=["upload","microphone"],
+                            type="filepath",
+                            label="Your Noisy Audio",
+                            interactive=True)
+    run_button   = gr.Button("Enhance", variant="primary")
+    enhanced_out = gr.Audio(label="Enhanced Audio", type="filepath", interactive=False)
+    spec_out     = gr.Plot(label="Spectrogram")
+
+    run_button.click(enhance, inputs=audio_in, outputs=[enhanced_out, spec_out])
+
+demo.queue().launch()