Update app.py

app.py

@@ -25,34 +25,6 @@ print("Loading DeepFilterNet model...", flush=True)
 df_model, df_state, _ = init_df()
 print("DeepFilterNet model ready!")
 
-
-def compress_dynamic_range(audio, sr, threshold_db=-28, ratio=4.0, makeup_gain_db=8, block_ms=20):
-    """
-    صدای آروم رو نسبت به صدای بلند بلندتر می‌کنه (یکسان‌سازی حجم صدا)،
-    تا صدای یه نفر که از میکروفون دوره به صدای کسی که میکروفون دستشه نزدیک‌تر بشه.
-    """
-    audio = audio.astype(np.float32)
-    block_size = max(1, int(sr * block_ms / 1000))
-    n_blocks = int(np.ceil(len(audio) / block_size))
-    pad_len = n_blocks * block_size - len(audio)
-    padded = np.pad(audio, (0, pad_len))
-    blocks = padded.reshape(n_blocks, block_size)
-
-    rms = np.sqrt(np.mean(blocks ** 2, axis=1) + 1e-10)
-    rms_db = 20 * np.log10(rms + 1e-10)
-
-    target_db = np.where(rms_db > threshold_db, threshold_db + (rms_db - threshold_db) / ratio, rms_db)
-    gain_db = target_db - rms_db + makeup_gain_db
-    gain = 10 ** (gain_db / 20)
-
-    kernel = np.ones(5, dtype=np.float32) / 5
-    gain_smooth = np.convolve(gain, kernel, mode="same")
-
-    gain_per_sample = np.repeat(gain_smooth, block_size)[:len(audio)].astype(np.float32)
-    compressed = audio * gain_per_sample
-    return np.clip(compressed, -1.0, 1.0).astype(np.float32)
-
-
 @app.get("/health")
 async def health():
     return {"status": "ok", "message": "SubFlow backend is running"}
@@ -76,12 +48,9 @@ async def transcribe(websocket: WebSocket):
         print(f"Step 3: saved to {tmp_path}")
 
         await websocket.send_json({"status": "cleaning"})
-        print("Step 3.5: leveling volume + cleaning audio with DeepFilterNet")
+        print("Step 3.5: cleaning audio with DeepFilterNet")
 
         audio_data, _ = load_audio(tmp_path, sr=df_state.sr())
-        audio_np = compress_dynamic_range(audio_data.numpy()[0], df_state.sr())
-        audio_data = torch.from_numpy(audio_np).unsqueeze(0)
-
         enhanced_audio = enhance(df_model, df_state, audio_data, atten_lim_db=20)
         enhanced_path = tmp_path.replace(".wav", "_enhanced.wav")
         save_audio(enhanced_path, enhanced_audio, df_state.sr())
@@ -98,6 +67,7 @@ async def transcribe(websocket: WebSocket):
             vad_parameters=dict(
                 min_silence_duration_ms=500,
                 speech_pad_ms=200,
+                threshold=0.3,
             ),
         )