update

cosyvoice/hifigan/f0_predictor.py

@@ -17,6 +17,7 @@ try:
     from torch.nn.utils.parametrizations import weight_norm
 except ImportError:
     from torch.nn.utils import weight_norm
+from cosyvoice.transformer.convolution import CausalConv1d
 
 
 class ConvRNNF0Predictor(nn.Module):
@@ -56,3 +57,47 @@ class ConvRNNF0Predictor(nn.Module):
         x = self.condnet(x)
         x = x.transpose(1, 2)
         return torch.abs(self.classifier(x).squeeze(-1))
+
+
+class CausalConvRNNF0Predictor(nn.Module):
+    def __init__(self,
+                 num_class: int = 1,
+                 in_channels: int = 80,
+                 cond_channels: int = 512
+                 ):
+        super().__init__()
+
+        self.num_class = num_class
+        self.condnet = nn.Sequential(
+            weight_norm(
+                CausalConv1d(in_channels, cond_channels, kernel_size=4, causal_type='right')
+            ),
+            nn.ELU(),
+            weight_norm(
+                CausalConv1d(cond_channels, cond_channels, kernel_size=3, causal_type='left')
+            ),
+            nn.ELU(),
+            weight_norm(
+                CausalConv1d(cond_channels, cond_channels, kernel_size=3, causal_type='left')
+            ),
+            nn.ELU(),
+            weight_norm(
+                CausalConv1d(cond_channels, cond_channels, kernel_size=3, causal_type='left')
+            ),
+            nn.ELU(),
+            weight_norm(
+                CausalConv1d(cond_channels, cond_channels, kernel_size=3, causal_type='left')
+            ),
+            nn.ELU(),
+        )
+        self.classifier = nn.Linear(in_features=cond_channels, out_features=self.num_class)
+
+    def forward(self, x: torch.Tensor, finalize: bool = True) -> torch.Tensor:
+        if finalize is True:
+            x = self.condnet[0](x)
+        else:
+            x = self.condnet[0](x[:, :, :-self.condnet[0].causal_padding], x[:, :, -self.condnet[0].causal_padding:])
+        for i in range(1, len(self.condnet)):
+            x = self.condnet[i](x)
+        x = x.transpose(1, 2)
+        return torch.abs(self.classifier(x).squeeze(-1))

cosyvoice/hifigan/generator.py

@@ -28,7 +28,7 @@ try:
 except ImportError:
     from torch.nn.utils import weight_norm
 from torch.distributions.uniform import Uniform
-
+from cosyvoice.transformer.convolution import CausalConv1d, CausalConv1dDownSample, CausalConv1dUpsample
 from cosyvoice.transformer.activation import Snake
 from cosyvoice.utils.common import get_padding
 from cosyvoice.utils.common import init_weights
@@ -50,8 +50,10 @@ class ResBlock(torch.nn.Module):
         channels: int = 512,
         kernel_size: int = 3,
         dilations: List[int] = [1, 3, 5],
+        causal: bool = False,
     ):
         super(ResBlock, self).__init__()
+        self.causal = causal
         self.convs1 = nn.ModuleList()
         self.convs2 = nn.ModuleList()
 
@@ -64,7 +66,14 @@ class ResBlock(torch.nn.Module):
                         kernel_size,
                         1,
                         dilation=dilation,
-                        padding=get_padding(kernel_size, dilation)
+                        padding=get_padding(kernel_size, dilation)) if causal is False else
+                    CausalConv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation,
+                        causal_type='left'
                     )
                 )
             )
@@ -76,7 +85,14 @@ class ResBlock(torch.nn.Module):
                         kernel_size,
                         1,
                         dilation=1,
-                        padding=get_padding(kernel_size, 1)
+                        padding=get_padding(kernel_size, 1)) if causal is False else
+                    CausalConv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        causal_type='left'
                     )
                 )
             )
@@ -171,58 +187,6 @@ class SineGen(torch.nn.Module):
         return sine_waves, uv, noise
 
 
-class SourceModuleHnNSF(torch.nn.Module):
-    """ SourceModule for hn-nsf
-    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
-                 add_noise_std=0.003, voiced_threshod=0)
-    sampling_rate: sampling_rate in Hz
-    harmonic_num: number of harmonic above F0 (default: 0)
-    sine_amp: amplitude of sine source signal (default: 0.1)
-    add_noise_std: std of additive Gaussian noise (default: 0.003)
-        note that amplitude of noise in unvoiced is decided
-        by sine_amp
-    voiced_threshold: threhold to set U/V given F0 (default: 0)
-    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
-    F0_sampled (batchsize, length, 1)
-    Sine_source (batchsize, length, 1)
-    noise_source (batchsize, length 1)
-    uv (batchsize, length, 1)
-    """
-
-    def __init__(self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1,
-                 add_noise_std=0.003, voiced_threshod=0):
-        super(SourceModuleHnNSF, self).__init__()
-
-        self.sine_amp = sine_amp
-        self.noise_std = add_noise_std
-
-        # to produce sine waveforms
-        self.l_sin_gen = SineGen(sampling_rate, harmonic_num,
-                                 sine_amp, add_noise_std, voiced_threshod)
-
-        # to merge source harmonics into a single excitation
-        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
-        self.l_tanh = torch.nn.Tanh()
-
-    def forward(self, x):
-        """
-        Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
-        F0_sampled (batchsize, length, 1)
-        Sine_source (batchsize, length, 1)
-        noise_source (batchsize, length 1)
-        """
-        # source for harmonic branch
-        with torch.no_grad():
-            sine_wavs, uv, _ = self.l_sin_gen(x.transpose(1, 2))
-            sine_wavs = sine_wavs.transpose(1, 2)
-            uv = uv.transpose(1, 2)
-        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
-
-        # source for noise branch, in the same shape as uv
-        noise = torch.randn_like(uv) * self.sine_amp / 3
-        return sine_merge, noise, uv
-
-
 class SineGen2(torch.nn.Module):
     """ Definition of sine generator
     SineGen(samp_rate, harmonic_num = 0,
@@ -242,7 +206,8 @@ class SineGen2(torch.nn.Module):
     def __init__(self, samp_rate, upsample_scale, harmonic_num=0,
                  sine_amp=0.1, noise_std=0.003,
                  voiced_threshold=0,
-                 flag_for_pulse=False):
+                 flag_for_pulse=False,
+                 causal=False):
         super(SineGen2, self).__init__()
         self.sine_amp = sine_amp
         self.noise_std = noise_std
@@ -252,6 +217,11 @@ class SineGen2(torch.nn.Module):
         self.voiced_threshold = voiced_threshold
         self.flag_for_pulse = flag_for_pulse
         self.upsample_scale = upsample_scale
+        self.causal = causal
+        if causal is True:
+            self.rand_ini = torch.rand(1, 9)
+            self.rand_ini[:, 0] = 0
+            self.sine_waves = torch.rand(1, 60 * 16000, 9)
 
     def _f02uv(self, f0):
         # generate uv signal
@@ -267,9 +237,12 @@ class SineGen2(torch.nn.Module):
         rad_values = (f0_values / self.sampling_rate) % 1
 
         # initial phase noise (no noise for fundamental component)
-        rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], device=f0_values.device)
-        rand_ini[:, 0] = 0
-        rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+        if self.training is False and self.causal is True:
+            rad_values[:, 0, :] = rad_values[:, 0, :] + self.rand_ini.to(rad_values.device)
+        else:
+            rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], device=f0_values.device)
+            rand_ini[:, 0] = 0
+            rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
 
         # instantanouse phase sine[t] = sin(2*pi \sum_i=1 ^{t} rad)
         if not self.flag_for_pulse:
@@ -279,7 +252,7 @@ class SineGen2(torch.nn.Module):
 
             phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
             phase = torch.nn.functional.interpolate(phase.transpose(1, 2) * self.upsample_scale,
-                                                    scale_factor=self.upsample_scale, mode="linear").transpose(1, 2)
+                                                    scale_factor=self.upsample_scale, mode="nearest" if self.causal is True else 'linear').transpose(1, 2)
             sines = torch.sin(phase)
         else:
             # If necessary, make sure that the first time step of every
@@ -331,7 +304,10 @@ class SineGen2(torch.nn.Module):
         #        std = self.sine_amp/3 -> max value ~ self.sine_amp
         # .       for voiced regions is self.noise_std
         noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
-        noise = noise_amp * torch.randn_like(sine_waves)
+        if self.training is False and self.causal is True:
+            noise = noise_amp * self.sine_waves[:, :sine_waves.shape[1]].to(sine_waves.device)
+        else:
+            noise = noise_amp * torch.randn_like(sine_waves)
 
         # first: set the unvoiced part to 0 by uv
         # then: additive noise
@@ -339,7 +315,7 @@ class SineGen2(torch.nn.Module):
         return sine_waves, uv, noise
 
 
-class SourceModuleHnNSF2(torch.nn.Module):
+class SourceModuleHnNSF(torch.nn.Module):
     """ SourceModule for hn-nsf
     SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
                  add_noise_std=0.003, voiced_threshod=0)
@@ -358,19 +334,26 @@ class SourceModuleHnNSF2(torch.nn.Module):
     """
 
     def __init__(self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1,
-                 add_noise_std=0.003, voiced_threshod=0):
-        super(SourceModuleHnNSF2, self).__init__()
+                 add_noise_std=0.003, voiced_threshod=0, sinegen_type='1', causal=False):
+        super(SourceModuleHnNSF, self).__init__()
 
         self.sine_amp = sine_amp
         self.noise_std = add_noise_std
 
         # to produce sine waveforms
-        self.l_sin_gen = SineGen2(sampling_rate, upsample_scale, harmonic_num,
-                                  sine_amp, add_noise_std, voiced_threshod)
+        if sinegen_type == '1':
+            self.l_sin_gen = SineGen(sampling_rate, harmonic_num,
+                                    sine_amp, add_noise_std, voiced_threshod)
+        else:
+            self.l_sin_gen = SineGen2(sampling_rate, upsample_scale, harmonic_num,
+                                    sine_amp, add_noise_std, voiced_threshod, causal=causal)
 
         # to merge source harmonics into a single excitation
         self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
         self.l_tanh = torch.nn.Tanh()
+        self.causal = causal
+        if causal is True:
+            self.uv = torch.rand(1, 60 * 24000, 1)
 
     def forward(self, x):
         """
@@ -385,7 +368,10 @@ class SourceModuleHnNSF2(torch.nn.Module):
         sine_merge = self.l_tanh(self.l_linear(sine_wavs))
 
         # source for noise branch, in the same shape as uv
-        noise = torch.randn_like(uv) * self.sine_amp / 3
+        if self.training is False and self.causal is True:
+            noise = self.uv[:, :uv.shape[1]] * self.sine_amp / 3
+        else:
+            noise = torch.randn_like(uv) * self.sine_amp / 3
         return sine_merge, noise, uv
 
 
@@ -425,15 +411,16 @@ class HiFTGenerator(nn.Module):
 
         self.num_kernels = len(resblock_kernel_sizes)
         self.num_upsamples = len(upsample_rates)
-        # NOTE in CosyVoice2, we use the original SourceModuleHnNSF implementation
-        this_SourceModuleHnNSF = SourceModuleHnNSF if self.sampling_rate == 22050 else SourceModuleHnNSF2
-        self.m_source = this_SourceModuleHnNSF(
+        # NOTE in CosyVoice2, we use the original SineGen implementation
+        self.m_source = SourceModuleHnNSF(
             sampling_rate=sampling_rate,
             upsample_scale=np.prod(upsample_rates) * istft_params["hop_len"],
             harmonic_num=nb_harmonics,
             sine_amp=nsf_alpha,
             add_noise_std=nsf_sigma,
-            voiced_threshod=nsf_voiced_threshold)
+            voiced_threshod=nsf_voiced_threshold,
+            sinegen_type='1' if self.sampling_rate == 22050 else '2',
+            causal=False)
         self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates) * istft_params["hop_len"])
 
         self.conv_pre = weight_norm(
@@ -580,3 +567,179 @@ class HiFTGenerator(nn.Module):
             s[:, :, :cache_source.shape[2]] = cache_source
         generated_speech = self.decode(x=speech_feat, s=s)
         return generated_speech, s
+
+
+class CausalHiFTGenerator(HiFTGenerator):
+    """
+    HiFTNet Generator: Neural Source Filter + ISTFTNet
+    https://arxiv.org/abs/2309.09493
+    """
+    def __init__(
+            self,
+            in_channels: int = 80,
+            base_channels: int = 512,
+            nb_harmonics: int = 8,
+            sampling_rate: int = 22050,
+            nsf_alpha: float = 0.1,
+            nsf_sigma: float = 0.003,
+            nsf_voiced_threshold: float = 10,
+            upsample_rates: List[int] = [8, 8],
+            upsample_kernel_sizes: List[int] = [16, 16],
+            istft_params: Dict[str, int] = {"n_fft": 16, "hop_len": 4},
+            resblock_kernel_sizes: List[int] = [3, 7, 11],
+            resblock_dilation_sizes: List[List[int]] = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+            source_resblock_kernel_sizes: List[int] = [7, 11],
+            source_resblock_dilation_sizes: List[List[int]] = [[1, 3, 5], [1, 3, 5]],
+            lrelu_slope: float = 0.1,
+            audio_limit: float = 0.99,
+            conv_pre_look_right: int = 4,
+            f0_predictor: torch.nn.Module = None,
+    ):
+        torch.nn.Module.__init__(self)
+
+        self.out_channels = 1
+        self.nb_harmonics = nb_harmonics
+        self.sampling_rate = sampling_rate
+        self.istft_params = istft_params
+        self.lrelu_slope = lrelu_slope
+        self.audio_limit = audio_limit
+
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sampling_rate,
+            upsample_scale=np.prod(upsample_rates) * istft_params["hop_len"],
+            harmonic_num=nb_harmonics,
+            sine_amp=nsf_alpha,
+            add_noise_std=nsf_sigma,
+            voiced_threshod=nsf_voiced_threshold,
+            sinegen_type='1' if self.sampling_rate == 22050 else '2',
+            causal=True)
+        self.upsample_rates = upsample_rates
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates) * istft_params["hop_len"])
+
+        self.conv_pre = weight_norm(
+            CausalConv1d(in_channels, base_channels, conv_pre_look_right + 1, 1, causal_type='right')
+        )
+
+        # Up
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    CausalConv1dUpsample(
+                        base_channels // (2**i),
+                        base_channels // (2**(i + 1)),
+                        k,
+                        u,
+                    )
+                )
+            )
+
+        # Down
+        self.source_downs = nn.ModuleList()
+        self.source_resblocks = nn.ModuleList()
+        downsample_rates = [1] + upsample_rates[::-1][:-1]
+        downsample_cum_rates = np.cumprod(downsample_rates)
+        for i, (u, k, d) in enumerate(zip(downsample_cum_rates[::-1], source_resblock_kernel_sizes, source_resblock_dilation_sizes)):
+            if u == 1:
+                self.source_downs.append(
+                    CausalConv1d(istft_params["n_fft"] + 2, base_channels // (2 ** (i + 1)), 1, 1, causal_type='left')
+                )
+            else:
+                self.source_downs.append(
+                    CausalConv1dDownSample(istft_params["n_fft"] + 2, base_channels // (2 ** (i + 1)), u * 2, u)
+                )
+
+            self.source_resblocks.append(
+                ResBlock(base_channels // (2 ** (i + 1)), k, d, causal=True)
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = base_channels // (2**(i + 1))
+            for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+                self.resblocks.append(ResBlock(ch, k, d, causal=True))
+
+        self.conv_post = weight_norm(CausalConv1d(ch, istft_params["n_fft"] + 2, 7, 1, causal_type='left'))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+        self.reflection_pad = nn.ReflectionPad1d((1, 0))
+        self.stft_window = torch.from_numpy(get_window("hann", istft_params["n_fft"], fftbins=True).astype(np.float32))
+        self.conv_pre_look_right = conv_pre_look_right
+        self.f0_predictor = f0_predictor
+
+    def decode(self, x: torch.Tensor, s: torch.Tensor = torch.zeros(1, 1, 0), finalize: bool = True) -> torch.Tensor:
+        s_stft_real, s_stft_imag = self._stft(s.squeeze(1))
+        if finalize is True:
+            x = self.conv_pre(x)
+        else:
+            x = self.conv_pre(x[:, :, :-self.conv_pre_look_right], x[:, :, -self.conv_pre_look_right:])
+            s_stft_real, s_stft_imag = s_stft_real[:, :, :-int(np.prod(self.upsample_rates) * self.conv_pre_look_right)], s_stft_imag[:, :, :-int(np.prod(self.upsample_rates) * self.conv_pre_look_right)]
+        s_stft = torch.cat([s_stft_real, s_stft_imag], dim=1)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, self.lrelu_slope)
+            x = self.ups[i](x)
+
+            if i == self.num_upsamples - 1:
+                x = self.reflection_pad(x)
+
+            # fusion
+            si = self.source_downs[i](s_stft)
+            si = self.source_resblocks[i](si)
+            x = x + si
+
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        magnitude = torch.exp(x[:, :self.istft_params["n_fft"] // 2 + 1, :])
+        phase = torch.sin(x[:, self.istft_params["n_fft"] // 2 + 1:, :])  # actually, sin is redundancy
+
+        x = self._istft(magnitude, phase)
+        if finalize is False:
+            x = x[:, :-int(np.prod(self.upsample_rates) * self.istft_params['hop_len'])]
+        x = torch.clamp(x, -self.audio_limit, self.audio_limit)
+        return x
+
+    @torch.inference_mode()
+    def inference(self, speech_feat: torch.Tensor, finalize: bool = True) -> torch.Tensor:
+        # mel->f0
+        self.f0_predictor.to('cpu')
+        f0 = self.f0_predictor(speech_feat.cpu(), finalize=finalize).to(speech_feat)
+        # f0->source
+        s = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
+        s, _, _ = self.m_source(s)
+        s = s.transpose(1, 2)
+        if finalize is True:
+            generated_speech = self.decode(x=speech_feat, s=s, finalize=finalize)
+        else:
+            generated_speech = self.decode(x=speech_feat[:, :, :-self.f0_predictor.condnet[0].causal_padding], s=s, finalize=finalize)
+        return generated_speech, s
+
+
+if __name__ == '__main__':
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+    from hyperpyyaml import load_hyperpyyaml
+    with open('./pretrained_models/CosyVoice3-0.5B/cosyvoice3.yaml', 'r') as f:
+        configs = load_hyperpyyaml(f, overrides={'llm': None, 'flow': None})
+    model = configs['hift']
+    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+    model.to(device)
+    model.eval()
+    max_len, chunk_size, context_size = 300, 30, 8
+    mel = torch.rand(1, 80, max_len)
+    pred_gt, _ = model.inference(mel)
+    for i in range(0, max_len, chunk_size):
+        finalize = True if i + chunk_size + context_size >= max_len else False
+        pred_chunk, _ = model.inference(mel[:, :, : i + chunk_size + context_size], finalize=finalize)
+        pred_chunk = pred_chunk[:, i * 480:]
+        print((pred_gt[:, i * 480:i * 480 + pred_chunk.shape[1]] - pred_chunk).abs().max().item())