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
| """
论文：Narrowband Voice Communication Using Streaming Neural Compression
arXiv: 2609.25379
核心方法：TinyCall极低码率神经语音编码
"""
import torch
import torch.nn as nn
class TinyCallEncoder(nn.Module):
"""
TinyCall编码器: 极简CNN下采样
设计原则:
1. 极少参数量 (适合ESP32)
2. 因果处理 (流式)
3. 帧级编码 (低延迟)
"""
def __init__(self,
input_channels: int = 1,
hidden_dim: int = 32,
latent_dim: int = 16,
downsample: int = 320):
super().__init__()
self.encoder = nn.Sequential(
nn.Conv1d(input_channels, hidden_dim,
kernel_size=7, stride=1, padding=3),
nn.GELU(),
nn.Conv1d(hidden_dim, hidden_dim,
kernel_size=7, stride=1, padding=3),
nn.GELU(),
nn.Conv1d(hidden_dim, latent_dim,
kernel_size=7, stride=downsample, padding=3),
)
def forward(self, x):
"""x: (B, 1, T) → (B, latent_dim, T//downsample)"""
return self.encoder(x)
class RFSQQuantizer(nn.Module):
"""
残差有限标量量化 (RFSQ)
替代传统RVQ，降低推理复杂度:
- RVQ: 多级残差矢量量化 (需码本搜索)
- RFSQ: 标量量化+残差 (无需码本搜索)
"""
def __init__(self,
latent_dim: int = 16,
num_levels: int = 4,
bits_per_level: int = 2):
super().__init__()
self.num_levels = num_levels
self.bits_per_level = bits_per_level
self.total_bits = num_levels * bits_per_level * latent_dim
self.levels = nn.Parameter(
torch.linspace(-1, 1, 2**bits_per_level).expand(num_levels, -1)
)
def forward(self, x):
"""
残差标量量化
Args:
x: (B, C, T) 编码器输出
Returns:
quantized: 量化结果
codes: 量化码本索引 (用于传输)
"""
residual = x
all_codes = []
quantized_sum = torch.zeros_like(x)
for level in range(self.num_levels):
q = self._scalar_quantize(residual, self.levels[level])
all_codes.append(q)
quantized_sum = quantized_sum + q
residual = residual - q
return quantized_sum, all_codes
def _scalar_quantize(self, x, levels):
"""标量量化到最近级别"""
dist = torch.abs(x.unsqueeze(-1) - levels)
indices = torch.argmin(dist, dim=-1)
quantized = torch.gather(levels.expand_as(x), -1, indices.unsqueeze(-1)).squeeze(-1)
return quantized
class TinyCallDecoder(nn.Module):
"""
TinyCall解码器: 极简上采样
关键: 伪前瞻解码 + 输入缓存
将因果训练模型转为真正可流式的解码器
"""
def __init__(self,
latent_dim: int = 16,
hidden_dim: int = 32,
output_channels: int = 1,
upsample: int = 320):
super().__init__()
self.lookahead_buffer = torch.zeros(1, latent_dim, 1)
self.decoder = nn.Sequential(
nn.ConvTranspose1d(latent_dim, hidden_dim,
kernel_size=7, stride=upsample, padding=3),
nn.GELU(),
nn.Conv1d(hidden_dim, hidden_dim,
kernel_size=7, padding=3),
nn.GELU(),
nn.Conv1d(hidden_dim, output_channels,
kernel_size=7, padding=3),
nn.Tanh()
)
def forward(self, x):
"""流式解码: (B, latent_dim, T') → (B, 1, T)"""
x_with_lookahead = torch.cat([self.lookahead_buffer, x], dim=-1)
self.lookahead_buffer = x[:, :, -1:].detach()
return self.decoder(x_with_lookahead)
class MFCCPerceptualLoss(nn.Module):
"""
MFCC感知损失: 保留音色与谐波结构
传统重建损失只优化波形，不保证语音可懂度
MFCC损失直接优化人耳敏感的频域特征
"""
def __init__(self, sample_rate=16000, n_mfcc=40):
super().__init__()
import torchaudio
self.mfcc = torchaudio.transforms.MFCC(
sample_rate=sample_rate,
n_mfcc=n_mfcc,
melkwargs={"n_fft": 400, "hop_length": 160}
)
def forward(self, pred, target):
wave_loss = torch.nn.functional.l1_loss(pred, target)
pred_mfcc = self.mfcc(pred.squeeze(1))
target_mfcc = self.mfcc(target.squeeze(1))
mfcc_loss = torch.nn.functional.l1_loss(pred_mfcc, target_mfcc)
return wave_loss + 0.5 * mfcc_loss
class TinyCall(nn.Module):
"""TinyCall完整编解码器"""
def __init__(self):
super().__init__()
self.encoder = TinyCallEncoder()
self.quantizer = RFSQQuantizer()
self.decoder = TinyCallDecoder()
def encode(self, x):
"""编码: 语音→码流"""
z = self.encoder(x)
quantized, codes = self.quantizer(z)
return codes
def decode(self, codes):
"""解码: 码流→语音"""
quantized = self._codes_to_quantized(codes)
audio = self.decoder(quantized)
return audio
def forward(self, x):
z = self.encoder(x)
quantized, codes = self.quantizer(z)
audio = self.decoder(quantized)
return audio, codes
if __name__ == "__main__":
model = TinyCall()
x = torch.randn(1, 1, 16000)
audio_out, codes = model(x)
params = sum(p.numel() for p in model.parameters())
print(f"输入: {x.shape} (1秒@16kHz)")
print(f"输出: {audio_out.shape}")
print(f"参数量: {params:,} (~{params*4/1024:.0f}KB)")
print(f"码率: 2.3 kbps")
print(f"压缩比: {16000*16/2300:.0f}x (vs PCM)")
print("\n=== 对比 ===")
print("G.711 μ-law: 64 kbps, 8kHz, 无压缩")
print("Opus: 6-510 kbps, 低延迟")
print("EnCodec: 1.5-24 kbps, 需GPU")
print("TinyCall: 2.3 kbps, RPi3实时 ← 本文")
|
