Initial audio denoiser gradio app

app.py

@@ -0,0 +1,121 @@
+import torch
+import librosa
+import soundfile as sf
+import numpy as np
+import tempfile
+import gradio as gr
+
+from denoiser import pretrained
+from denoiser.dsp import convert_audio
+from pydub import AudioSegment, silence
+from tqdm import tqdm
+
+
+# -----------------------------
+# Load model ONCE (important!)
+# -----------------------------
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model = pretrained.dns64().to(device)
+
+
+# -----------------------------
+# Silence trimming helpers
+# -----------------------------
+def safe_append(base, chunk, crossfade_ms=30):
+    if len(base) > 0 and len(chunk) > 0:
+        safe_crossfade = min(crossfade_ms, len(base), len(chunk))
+        if safe_crossfade > 0:
+            return base.append(chunk, crossfade=safe_crossfade)
+    return base + chunk
+
+
+def shorten_silences(audio, silence_thresh=-50, crossfade_ms=30):
+    silent_ranges = silence.detect_silence(
+        audio,
+        min_silence_len=400,
+        silence_thresh=silence_thresh
+    )
+
+    output = AudioSegment.silent(duration=0)
+    prev_end = 0
+
+    for start, end in silent_ranges:
+        chunk = audio[prev_end:start]
+        output = safe_append(output, chunk, crossfade_ms)
+
+        silence_len = end - start
+        if silence_len < 500:
+            keep = silence_len
+        elif silence_len <= 1500:
+            keep = 500
+        elif silence_len <= 2500:
+            keep = 1000
+        else:
+            keep = 1500
+
+        output = safe_append(
+            output,
+            AudioSegment.silent(duration=keep),
+            crossfade_ms
+        )
+        prev_end = end
+
+    output = safe_append(output, audio[prev_end:], crossfade_ms)
+    return output
+
+
+# -----------------------------
+# Main processing function
+# -----------------------------
+def denoise_audio(audio_file, trim_silence):
+    wav, sr = librosa.load(audio_file, sr=16000)
+
+    chunk_size = 16000 * 10
+    denoised_chunks = []
+
+    for i in range(0, len(wav), chunk_size):
+        chunk = wav[i:i + chunk_size]
+        wav_tensor = torch.tensor(chunk).unsqueeze(0).to(device)
+        wav_tensor = convert_audio(
+            wav_tensor, sr, model.sample_rate, model.chin
+        )
+
+        with torch.no_grad():
+            denoised = model(wav_tensor)[0]
+
+        denoised_chunks.append(
+            denoised.squeeze().cpu().numpy()
+        )
+
+    denoised_np = np.concatenate(denoised_chunks)
+
+    tmp_wav = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+    sf.write(tmp_wav.name, denoised_np, model.sample_rate)
+
+    if trim_silence:
+        audio = AudioSegment.from_file(tmp_wav.name, format="wav")
+        processed = shorten_silences(audio)
+        final_file = tempfile.NamedTemporaryFile(
+            suffix="_final.wav", delete=False
+        )
+        processed.export(final_file.name, format="wav")
+        return final_file.name
+
+    return tmp_wav.name
+
+
+# -----------------------------
+# Gradio UI
+# -----------------------------
+demo = gr.Interface(
+    fn=denoise_audio,
+    inputs=[
+        gr.Audio(type="filepath", label="Upload Audio"),
+        gr.Checkbox(label="Trim silence after denoising", value=True)
+    ],
+    outputs=gr.Audio(label="Denoised Output"),
+    title="🎧 Audio Denoiser (Demucs DNS64)",
+    description="Upload an audio file, optionally trim silences, and get clean audio."
+)
+
+demo.launch()

requirements.txt

@@ -0,0 +1,9 @@
+git+https://github.com/facebookresearch/denoiser.git
+torch
+torchaudio
+librosa
+soundfile
+pydub
+tqdm
+gradio
+numpy