Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -5,84 +5,133 @@ import os
|
|
5 |
from pydub import AudioSegment
|
6 |
import tempfile
|
7 |
from speechbrain.pretrained.separation import SepformerSeparation
|
|
|
|
|
|
|
|
|
8 |
|
9 |
-
class
|
10 |
def __init__(self):
|
11 |
-
# Initialize the
|
12 |
self.model = SepformerSeparation.from_hparams(
|
13 |
source="speechbrain/sepformer-dns4-16k-enhancement",
|
14 |
savedir='pretrained_models/sepformer-dns4-16k-enhancement'
|
15 |
)
|
16 |
|
17 |
-
#
|
18 |
-
|
19 |
-
|
20 |
-
def convert_audio_to_wav(self, input_path):
|
21 |
-
"""
|
22 |
-
Convert any audio format to WAV with proper settings
|
23 |
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
try:
|
31 |
-
#
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
# Load audio using pydub (supports multiple formats)
|
36 |
-
audio = AudioSegment.from_file(input_path)
|
37 |
-
|
38 |
-
# Convert to mono if stereo
|
39 |
-
if audio.channels > 1:
|
40 |
-
audio = audio.set_channels(1)
|
41 |
|
42 |
-
#
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
parameters=[
|
47 |
-
'-ar', '16000', # Set sample rate to 16kHz
|
48 |
-
'-ac', '1' # Set channels to mono
|
49 |
-
]
|
50 |
-
)
|
51 |
|
52 |
-
return
|
53 |
|
54 |
except Exception as e:
|
55 |
-
|
56 |
-
|
57 |
-
|
|
|
58 |
"""
|
59 |
-
Process
|
60 |
-
|
61 |
-
Args:
|
62 |
-
audio_path (str): Path to the input audio file
|
63 |
-
|
64 |
-
Returns:
|
65 |
-
str: Path to the enhanced audio file
|
66 |
"""
|
67 |
try:
|
68 |
-
# Convert input audio to proper
|
69 |
-
|
|
|
|
|
70 |
|
71 |
-
#
|
72 |
-
|
|
|
73 |
|
74 |
-
#
|
75 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
|
77 |
-
#
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
)
|
83 |
|
84 |
-
#
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
|
87 |
return output_path
|
88 |
|
@@ -91,11 +140,11 @@ class AudioDenoiser:
|
|
91 |
|
92 |
def create_gradio_interface():
|
93 |
# Initialize the denoiser
|
94 |
-
denoiser =
|
95 |
|
96 |
# Create the Gradio interface
|
97 |
interface = gr.Interface(
|
98 |
-
fn=denoiser.
|
99 |
inputs=gr.Audio(
|
100 |
type="filepath",
|
101 |
label="Upload Noisy Audio"
|
@@ -104,21 +153,10 @@ def create_gradio_interface():
|
|
104 |
label="Enhanced Audio",
|
105 |
type="filepath"
|
106 |
),
|
107 |
-
title="Audio Denoising using SepFormer",
|
108 |
description="""
|
109 |
-
|
110 |
-
|
111 |
-
""",
|
112 |
-
article="""
|
113 |
-
Supported audio formats:
|
114 |
-
- MP3
|
115 |
-
- WAV
|
116 |
-
- OGG
|
117 |
-
- FLAC
|
118 |
-
- M4A
|
119 |
-
and more...
|
120 |
-
|
121 |
-
The audio will automatically be converted to the correct format for processing.
|
122 |
"""
|
123 |
)
|
124 |
|
|
|
5 |
from pydub import AudioSegment
|
6 |
import tempfile
|
7 |
from speechbrain.pretrained.separation import SepformerSeparation
|
8 |
+
import numpy as np
|
9 |
+
import threading
|
10 |
+
from queue import Queue
|
11 |
+
import time
|
12 |
|
13 |
+
class RealtimeAudioDenoiser:
|
14 |
def __init__(self):
|
15 |
+
# Initialize the model
|
16 |
self.model = SepformerSeparation.from_hparams(
|
17 |
source="speechbrain/sepformer-dns4-16k-enhancement",
|
18 |
savedir='pretrained_models/sepformer-dns4-16k-enhancement'
|
19 |
)
|
20 |
|
21 |
+
# Move model to GPU if available
|
22 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
23 |
+
self.model.to(self.device)
|
|
|
|
|
|
|
24 |
|
25 |
+
# Enable inference mode for better performance
|
26 |
+
self.model.eval()
|
27 |
+
torch.set_grad_enabled(False)
|
28 |
+
|
29 |
+
# Set chunk size for streaming (500ms chunks)
|
30 |
+
self.chunk_duration = 0.5 # seconds
|
31 |
+
self.sample_rate = 16000
|
32 |
+
self.chunk_size = int(self.sample_rate * self.chunk_duration)
|
33 |
+
|
34 |
+
# Initialize processing queue and buffer
|
35 |
+
self.processing_queue = Queue()
|
36 |
+
self.output_buffer = Queue()
|
37 |
+
self.is_processing = False
|
38 |
+
|
39 |
+
# Start processing thread
|
40 |
+
self.processing_thread = threading.Thread(target=self._process_queue)
|
41 |
+
self.processing_thread.daemon = True
|
42 |
+
self.processing_thread.start()
|
43 |
+
|
44 |
+
# Create output directory
|
45 |
+
os.makedirs("enhanced_audio", exist_ok=True)
|
46 |
+
|
47 |
+
def _optimize_model(self):
|
48 |
+
"""Optimize model for inference"""
|
49 |
+
if self.device.type == 'cuda':
|
50 |
+
# Use mixed precision for faster processing
|
51 |
+
self.model = torch.quantization.quantize_dynamic(
|
52 |
+
self.model, {torch.nn.Linear}, dtype=torch.qint8
|
53 |
+
)
|
54 |
+
torch.backends.cudnn.benchmark = True
|
55 |
+
|
56 |
+
def _process_queue(self):
|
57 |
+
"""Background thread for processing audio chunks"""
|
58 |
+
while True:
|
59 |
+
if not self.processing_queue.empty():
|
60 |
+
chunk = self.processing_queue.get()
|
61 |
+
if chunk is None:
|
62 |
+
continue
|
63 |
+
|
64 |
+
# Process audio chunk
|
65 |
+
enhanced_chunk = self._enhance_chunk(chunk)
|
66 |
+
self.output_buffer.put(enhanced_chunk)
|
67 |
+
else:
|
68 |
+
time.sleep(0.01) # Small delay to prevent CPU overuse
|
69 |
+
|
70 |
+
def _enhance_chunk(self, audio_chunk):
|
71 |
+
"""Process a single chunk of audio"""
|
72 |
try:
|
73 |
+
# Convert to tensor and move to device
|
74 |
+
chunk_tensor = torch.FloatTensor(audio_chunk).to(self.device)
|
75 |
+
chunk_tensor = chunk_tensor.unsqueeze(0) # Add batch dimension
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
|
77 |
+
# Process with model
|
78 |
+
with torch.inference_mode():
|
79 |
+
enhanced = self.model.separate_batch(chunk_tensor)
|
80 |
+
enhanced = enhanced.squeeze(0).cpu().numpy()
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
+
return enhanced
|
83 |
|
84 |
except Exception as e:
|
85 |
+
print(f"Error processing chunk: {str(e)}")
|
86 |
+
return audio_chunk
|
87 |
+
|
88 |
+
def process_stream(self, audio_path):
|
89 |
"""
|
90 |
+
Process audio in streaming fashion
|
|
|
|
|
|
|
|
|
|
|
|
|
91 |
"""
|
92 |
try:
|
93 |
+
# Convert input audio to proper format
|
94 |
+
audio = AudioSegment.from_file(audio_path)
|
95 |
+
audio = audio.set_frame_rate(self.sample_rate)
|
96 |
+
audio = audio.set_channels(1)
|
97 |
|
98 |
+
# Convert to numpy array
|
99 |
+
samples = np.array(audio.get_array_of_samples(), dtype=np.float32)
|
100 |
+
samples = samples / np.max(np.abs(samples)) # Normalize
|
101 |
|
102 |
+
# Process in chunks
|
103 |
+
enhanced_chunks = []
|
104 |
+
for i in range(0, len(samples), self.chunk_size):
|
105 |
+
chunk = samples[i:i + self.chunk_size]
|
106 |
+
|
107 |
+
# Pad last chunk if necessary
|
108 |
+
if len(chunk) < self.chunk_size:
|
109 |
+
chunk = np.pad(chunk, (0, self.chunk_size - len(chunk)))
|
110 |
+
|
111 |
+
# Add to processing queue
|
112 |
+
self.processing_queue.put(chunk)
|
113 |
|
114 |
+
# Wait for all chunks to be processed
|
115 |
+
while self.processing_queue.qsize() > 0 or self.output_buffer.qsize() > 0:
|
116 |
+
if not self.output_buffer.empty():
|
117 |
+
enhanced_chunks.append(self.output_buffer.get())
|
118 |
+
time.sleep(0.01)
|
|
|
119 |
|
120 |
+
# Combine chunks
|
121 |
+
enhanced_audio = np.concatenate(enhanced_chunks)
|
122 |
+
|
123 |
+
# Save enhanced audio
|
124 |
+
output_path = os.path.join("enhanced_audio", "enhanced_realtime.wav")
|
125 |
+
enhanced_audio = enhanced_audio * 32767 # Convert to int16 range
|
126 |
+
enhanced_audio = enhanced_audio.astype(np.int16)
|
127 |
+
|
128 |
+
with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as f:
|
129 |
+
torchaudio.save(
|
130 |
+
f.name,
|
131 |
+
torch.tensor(enhanced_audio).unsqueeze(0),
|
132 |
+
self.sample_rate
|
133 |
+
)
|
134 |
+
os.replace(f.name, output_path)
|
135 |
|
136 |
return output_path
|
137 |
|
|
|
140 |
|
141 |
def create_gradio_interface():
|
142 |
# Initialize the denoiser
|
143 |
+
denoiser = RealtimeAudioDenoiser()
|
144 |
|
145 |
# Create the Gradio interface
|
146 |
interface = gr.Interface(
|
147 |
+
fn=denoiser.process_stream,
|
148 |
inputs=gr.Audio(
|
149 |
type="filepath",
|
150 |
label="Upload Noisy Audio"
|
|
|
153 |
label="Enhanced Audio",
|
154 |
type="filepath"
|
155 |
),
|
156 |
+
title="Real-time Audio Denoising using SepFormer",
|
157 |
description="""
|
158 |
+
Optimized for real-time processing with low latency.
|
159 |
+
Processes audio in 500ms chunks for streaming applications.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
160 |
"""
|
161 |
)
|
162 |
|