Spaces:
Running
Running
File size: 9,135 Bytes
3a19dc4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 |
"""
DTLN Model Optimized for Alif E7 Ethos-U55 NPU
Lightweight voice denoising with 8-bit quantization support
"""
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.layers import (
LSTM, Dense, Input, Multiply, Lambda, Concatenate
)
from tensorflow.keras.models import Model
import numpy as np
class DTLN_Ethos_U55:
"""
Dual-signal Transformation LSTM Network optimized for Ethos-U55
Key optimizations:
- Reduced parameters for <1MB model size
- 8-bit quantization aware architecture
- Stateful inference for real-time processing
- Memory-efficient for DTCM constraints
"""
def __init__(
self,
frame_len=512,
frame_shift=128,
lstm_units=128,
sampling_rate=16000,
use_stft=True
):
"""
Args:
frame_len: Length of audio frame (default 512 = 32ms @ 16kHz)
frame_shift: Frame hop size (default 128 = 8ms @ 16kHz)
lstm_units: Number of LSTM units (reduced to 128 for NPU)
sampling_rate: Audio sampling rate in Hz
use_stft: Use STFT domain processing
"""
self.frame_len = frame_len
self.frame_shift = frame_shift
self.lstm_units = lstm_units
self.sampling_rate = sampling_rate
self.use_stft = use_stft
# Frequency bins for STFT
self.freq_bins = frame_len // 2 + 1
def stft_layer(self, x):
"""Custom STFT layer using tf.signal"""
stft = tf.signal.stft(
x,
frame_length=self.frame_len,
frame_step=self.frame_shift,
fft_length=self.frame_len
)
mag = tf.abs(stft)
phase = tf.math.angle(stft)
return mag, phase
def istft_layer(self, mag, phase):
"""Custom inverse STFT layer"""
complex_spec = tf.cast(mag, tf.complex64) * tf.exp(
1j * tf.cast(phase, tf.complex64)
)
signal = tf.signal.inverse_stft(
complex_spec,
frame_length=self.frame_len,
frame_step=self.frame_shift,
fft_length=self.frame_len
)
return signal
def build_model(self, training=True):
"""
Build the full DTLN model for training
Args:
training: If True, builds training model. If False, builds inference model.
Returns:
Keras Model
"""
# Input: raw waveform
input_audio = Input(shape=(None,), name='input_audio')
# Reshape for processing
audio_reshaped = Lambda(
lambda x: tf.expand_dims(x, -1)
)(input_audio)
# STFT transformation
mag, phase = Lambda(
lambda x: self.stft_layer(x),
name='stft'
)(input_audio)
# === First Processing Stage ===
# Process magnitude spectrum
lstm_1 = LSTM(
self.lstm_units,
return_sequences=True,
name='lstm_1'
)(mag)
# Estimate magnitude mask
mask_1 = Dense(
self.freq_bins,
activation='sigmoid',
name='mask_1'
)(lstm_1)
# Apply mask
enhanced_mag_1 = Multiply(name='apply_mask_1')([mag, mask_1])
# === Second Processing Stage ===
lstm_2 = LSTM(
self.lstm_units,
return_sequences=True,
name='lstm_2'
)(enhanced_mag_1)
# Second mask estimation
mask_2 = Dense(
self.freq_bins,
activation='sigmoid',
name='mask_2'
)(lstm_2)
# Apply second mask
enhanced_mag = Multiply(name='apply_mask_2')([enhanced_mag_1, mask_2])
# Inverse STFT
enhanced_audio = Lambda(
lambda x: self.istft_layer(x[0], x[1]),
name='istft'
)([enhanced_mag, phase])
# Build model
model = Model(
inputs=input_audio,
outputs=enhanced_audio,
name='DTLN_Ethos_U55'
)
return model
def build_stateful_model(self, batch_size=1):
"""
Build stateful model for frame-by-frame inference
This is more memory efficient for real-time processing
Returns:
Two models (stage1, stage2) for sequential processing
"""
# === Stage 1 Model ===
# Inputs
mag_input = Input(
batch_shape=(batch_size, 1, self.freq_bins),
name='magnitude_input'
)
state_h_1 = Input(
batch_shape=(batch_size, self.lstm_units),
name='lstm_1_state_h'
)
state_c_1 = Input(
batch_shape=(batch_size, self.lstm_units),
name='lstm_1_state_c'
)
# LSTM with state
lstm_1 = LSTM(
self.lstm_units,
return_sequences=True,
return_state=True,
stateful=False,
name='lstm_1'
)
lstm_out_1, state_h_1_out, state_c_1_out = lstm_1(
mag_input,
initial_state=[state_h_1, state_c_1]
)
# Mask estimation
mask_1 = Dense(
self.freq_bins,
activation='sigmoid',
name='mask_1'
)(lstm_out_1)
# Apply mask
enhanced_mag_1 = Multiply()([mag_input, mask_1])
model_1 = Model(
inputs=[mag_input, state_h_1, state_c_1],
outputs=[enhanced_mag_1, state_h_1_out, state_c_1_out],
name='DTLN_Stage1'
)
# === Stage 2 Model ===
mag_input_2 = Input(
batch_shape=(batch_size, 1, self.freq_bins),
name='enhanced_magnitude_input'
)
state_h_2 = Input(
batch_shape=(batch_size, self.lstm_units),
name='lstm_2_state_h'
)
state_c_2 = Input(
batch_shape=(batch_size, self.lstm_units),
name='lstm_2_state_c'
)
# LSTM with state
lstm_2 = LSTM(
self.lstm_units,
return_sequences=True,
return_state=True,
stateful=False,
name='lstm_2'
)
lstm_out_2, state_h_2_out, state_c_2_out = lstm_2(
mag_input_2,
initial_state=[state_h_2, state_c_2]
)
# Final mask
mask_2 = Dense(
self.freq_bins,
activation='sigmoid',
name='mask_2'
)(lstm_out_2)
# Apply mask
enhanced_mag = Multiply()([mag_input_2, mask_2])
model_2 = Model(
inputs=[mag_input_2, state_h_2, state_c_2],
outputs=[enhanced_mag, state_h_2_out, state_c_2_out],
name='DTLN_Stage2'
)
return model_1, model_2
def get_model_summary(self):
"""Print model architecture and parameter count"""
model = self.build_model()
model.summary()
total_params = model.count_params()
print(f"\nTotal parameters: {total_params:,}")
print(f"Estimated model size (FP32): {total_params * 4 / 1024:.2f} KB")
print(f"Estimated model size (INT8): {total_params / 1024:.2f} KB")
return model
def create_lightweight_model(target_size_kb=100):
"""
Factory function to create a lightweight model that fits in target size
Args:
target_size_kb: Target model size in KB
Returns:
DTLN_Ethos_U55 instance configured for target size
"""
# Estimate LSTM units for target size
# Rough estimate: each LSTM unit adds ~2KB for INT8
estimated_units = int((target_size_kb * 0.8) / 2)
estimated_units = min(max(estimated_units, 64), 256) # Clamp to 64-256
print(f"Creating model with {estimated_units} LSTM units")
print(f"Target size: {target_size_kb} KB")
model = DTLN_Ethos_U55(
frame_len=512,
frame_shift=128,
lstm_units=estimated_units,
sampling_rate=16000
)
return model
if __name__ == "__main__":
# Example usage
print("Creating DTLN model for Alif E7 Ethos-U55...")
# Create model
dtln = DTLN_Ethos_U55(
frame_len=512,
frame_shift=128,
lstm_units=128
)
# Get model summary
model = dtln.get_model_summary()
# Build stateful models for inference
print("\n" + "="*50)
print("Building stateful models for real-time inference...")
stage1, stage2 = dtln.build_stateful_model()
print("\nStage 1:")
stage1.summary()
print("\nStage 2:")
stage2.summary()
print("\n✓ Model creation successful!")
print("Next steps:")
print("1. Train the model with quantization-aware training")
print("2. Convert to TensorFlow Lite INT8 format")
print("3. Use Vela compiler to optimize for Ethos-U55")
|