Add support for hifigan and streaming

TTS/tts/layers/xtts/gpt.py

@@ -172,7 +172,7 @@ class GPT(nn.Module):
             "heads": list(self.text_head.parameters()) + list(self.mel_head.parameters()),
         }
 
-    def init_gpt_for_inference(self, kv_cache=True):
+    def init_gpt_for_inference(self, kv_cache=True, use_deepspeed=False):
         seq_length = self.max_prompt_tokens + self.max_mel_tokens + self.max_text_tokens + 1
         gpt_config = GPT2Config(
             vocab_size=self.max_mel_tokens,
@@ -195,6 +195,17 @@ class GPT(nn.Module):
         )
         self.gpt.wte = self.mel_embedding
 
+        if use_deepspeed:
+            import deepspeed
+            self.ds_engine = deepspeed.init_inference(
+                model=self.gpt_inference.half(),  # Transformers models
+                mp_size=1,  # Number of GPU
+                dtype=torch.float32,  # desired data type of output
+                replace_method="auto",  # Lets DS autmatically identify the layer to replace
+                replace_with_kernel_inject=True,  # replace the model with the kernel injector
+            )
+            self.gpt_inference = self.ds_engine.module.eval()
+
     def set_inputs_and_targets(self, input, start_token, stop_token):
         inp = F.pad(input, (1, 0), value=start_token)
         tar = F.pad(input, (0, 1), value=stop_token)
@@ -543,3 +554,14 @@ class GPT(nn.Module):
         if "return_dict_in_generate" in hf_generate_kwargs:
             return gen.sequences[:, gpt_inputs.shape[1] :], gen
         return gen[:, gpt_inputs.shape[1] :]
+
+    def get_generator(self, fake_inputs, **hf_generate_kwargs):
+        return self.gpt_inference.generate_stream(
+            fake_inputs,
+            bos_token_id=self.start_audio_token,
+            pad_token_id=self.stop_audio_token,
+            eos_token_id=self.stop_audio_token,
+            max_length=self.max_mel_tokens * 2 + self.max_prompt_tokens + self.max_text_tokens,
+            do_stream=True,
+            **hf_generate_kwargs,
+        )

TTS/tts/layers/xtts/hifigan_decoder.py

@@ -0,0 +1,742 @@
+import torch
+from torch import nn
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, weight_norm
+import torchaudio
+
+from TTS.utils.io import load_fsspec
+
+
+LRELU_SLOPE = 0.1
+
+
+def get_padding(k, d):
+    return int((k * d - d) / 2)
+
+
+class ResBlock1(torch.nn.Module):
+    """Residual Block Type 1. It has 3 convolutional layers in each convolutional block.
+
+    Network::
+
+        x -> lrelu -> conv1_1 -> conv1_2 -> conv1_3 -> z -> lrelu -> conv2_1 -> conv2_2 -> conv2_3 -> o -> + -> o
+        |--------------------------------------------------------------------------------------------------|
+
+
+    Args:
+        channels (int): number of hidden channels for the convolutional layers.
+        kernel_size (int): size of the convolution filter in each layer.
+        dilations (list): list of dilation value for each conv layer in a block.
+    """
+
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super().__init__()
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+
+    def forward(self, x):
+        """
+        Args:
+            x (Tensor): input tensor.
+        Returns:
+            Tensor: output tensor.
+        Shapes:
+            x: [B, C, T]
+        """
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(torch.nn.Module):
+    """Residual Block Type 2. It has 1 convolutional layers in each convolutional block.
+
+    Network::
+
+        x -> lrelu -> conv1-> -> z -> lrelu -> conv2-> o -> + -> o
+        |---------------------------------------------------|
+
+
+    Args:
+        channels (int): number of hidden channels for the convolutional layers.
+        kernel_size (int): size of the convolution filter in each layer.
+        dilations (list): list of dilation value for each conv layer in a block.
+    """
+
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super().__init__()
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+
+    def forward(self, x):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class HifiganGenerator(torch.nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        resblock_type,
+        resblock_dilation_sizes,
+        resblock_kernel_sizes,
+        upsample_kernel_sizes,
+        upsample_initial_channel,
+        upsample_factors,
+        inference_padding=5,
+        cond_channels=0,
+        conv_pre_weight_norm=True,
+        conv_post_weight_norm=True,
+        conv_post_bias=True,
+        cond_in_each_up_layer=False,
+    ):
+        r"""HiFiGAN Generator with Multi-Receptive Field Fusion (MRF)
+
+        Network:
+            x -> lrelu -> upsampling_layer -> resblock1_k1x1 -> z1 -> + -> z_sum / #resblocks -> lrelu -> conv_post_7x1 -> tanh -> o
+                                                 ..          -> zI ---|
+                                              resblockN_kNx1 -> zN ---'
+
+        Args:
+            in_channels (int): number of input tensor channels.
+            out_channels (int): number of output tensor channels.
+            resblock_type (str): type of the `ResBlock`. '1' or '2'.
+            resblock_dilation_sizes (List[List[int]]): list of dilation values in each layer of a `ResBlock`.
+            resblock_kernel_sizes (List[int]): list of kernel sizes for each `ResBlock`.
+            upsample_kernel_sizes (List[int]): list of kernel sizes for each transposed convolution.
+            upsample_initial_channel (int): number of channels for the first upsampling layer. This is divided by 2
+                for each consecutive upsampling layer.
+            upsample_factors (List[int]): upsampling factors (stride) for each upsampling layer.
+            inference_padding (int): constant padding applied to the input at inference time. Defaults to 5.
+        """
+        super().__init__()
+        self.inference_padding = inference_padding
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_factors)
+        self.cond_in_each_up_layer = cond_in_each_up_layer
+
+        # initial upsampling layers
+        self.conv_pre = weight_norm(
+            Conv1d(in_channels, upsample_initial_channel, 7, 1, padding=3)
+        )
+        resblock = ResBlock1 if resblock_type == "1" else ResBlock2
+        # upsampling layers
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_factors, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+        # MRF blocks
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for _, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+        # post convolution layer
+        self.conv_post = weight_norm(
+            Conv1d(ch, out_channels, 7, 1, padding=3, bias=conv_post_bias)
+        )
+        if cond_channels > 0:
+            self.cond_layer = nn.Conv1d(cond_channels, upsample_initial_channel, 1)
+
+        if not conv_pre_weight_norm:
+            remove_weight_norm(self.conv_pre)
+
+        if not conv_post_weight_norm:
+            remove_weight_norm(self.conv_post)
+
+        if self.cond_in_each_up_layer:
+            self.conds = nn.ModuleList()
+            for i in range(len(self.ups)):
+                ch = upsample_initial_channel // (2 ** (i + 1))
+                self.conds.append(nn.Conv1d(cond_channels, ch, 1))
+
+    def forward(self, x, g=None):
+        """
+        Args:
+            x (Tensor): feature input tensor.
+            g (Tensor): global conditioning input tensor.
+
+        Returns:
+            Tensor: output waveform.
+
+        Shapes:
+            x: [B, C, T]
+            Tensor: [B, 1, T]
+        """
+        o = self.conv_pre(x)
+        if hasattr(self, "cond_layer"):
+            o = o + self.cond_layer(g)
+        for i in range(self.num_upsamples):
+            o = F.leaky_relu(o, LRELU_SLOPE)
+            o = self.ups[i](o)
+
+            if self.cond_in_each_up_layer:
+                o = o + self.conds[i](g)
+
+            z_sum = None
+            for j in range(self.num_kernels):
+                if z_sum is None:
+                    z_sum = self.resblocks[i * self.num_kernels + j](o)
+                else:
+                    z_sum += self.resblocks[i * self.num_kernels + j](o)
+            o = z_sum / self.num_kernels
+        o = F.leaky_relu(o)
+        o = self.conv_post(o)
+        o = torch.tanh(o)
+        return o
+
+    @torch.no_grad()
+    def inference(self, c):
+        """
+        Args:
+            x (Tensor): conditioning input tensor.
+
+        Returns:
+            Tensor: output waveform.
+
+        Shapes:
+            x: [B, C, T]
+            Tensor: [B, 1, T]
+        """
+        c = c.to(self.conv_pre.weight.device)
+        c = torch.nn.functional.pad(
+            c, (self.inference_padding, self.inference_padding), "replicate"
+        )
+        return self.forward(c)
+
+    def remove_weight_norm(self):
+        print("Removing weight norm...")
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+    def load_checkpoint(
+        self, config, checkpoint_path, eval=False, cache=False
+    ):  # pylint: disable=unused-argument, redefined-builtin
+        state = torch.load(checkpoint_path, map_location=torch.device("cpu"))
+        self.load_state_dict(state["model"])
+        if eval:
+            self.eval()
+            assert not self.training
+            self.remove_weight_norm()
+
+class SELayer(nn.Module):
+    def __init__(self, channel, reduction=8):
+        super(SELayer, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.fc = nn.Sequential(
+            nn.Linear(channel, channel // reduction),
+            nn.ReLU(inplace=True),
+            nn.Linear(channel // reduction, channel),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, x):
+        b, c, _, _ = x.size()
+        y = self.avg_pool(x).view(b, c)
+        y = self.fc(y).view(b, c, 1, 1)
+        return x * y
+
+
+class SEBasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None, reduction=8):
+        super(SEBasicBlock, self).__init__()
+        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.relu = nn.ReLU(inplace=True)
+        self.se = SELayer(planes, reduction)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.relu(out)
+        out = self.bn1(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.se(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+        return out
+
+
+def set_init_dict(model_dict, checkpoint_state, c):
+    # Partial initialization: if there is a mismatch with new and old layer, it is skipped.
+    for k, v in checkpoint_state.items():
+        if k not in model_dict:
+            print(" | > Layer missing in the model definition: {}".format(k))
+    # 1. filter out unnecessary keys
+    pretrained_dict = {k: v for k, v in checkpoint_state.items() if k in model_dict}
+    # 2. filter out different size layers
+    pretrained_dict = {k: v for k, v in pretrained_dict.items() if v.numel() == model_dict[k].numel()}
+    # 3. skip reinit layers
+    if c.has("reinit_layers") and c.reinit_layers is not None:
+        for reinit_layer_name in c.reinit_layers:
+            pretrained_dict = {k: v for k, v in pretrained_dict.items() if reinit_layer_name not in k}
+    # 4. overwrite entries in the existing state dict
+    model_dict.update(pretrained_dict)
+    print(" | > {} / {} layers are restored.".format(len(pretrained_dict), len(model_dict)))
+    return model_dict
+
+
+class PreEmphasis(nn.Module):
+    def __init__(self, coefficient=0.97):
+        super().__init__()
+        self.coefficient = coefficient
+        self.register_buffer("filter", torch.FloatTensor([-self.coefficient, 1.0]).unsqueeze(0).unsqueeze(0))
+
+    def forward(self, x):
+        assert len(x.size()) == 2
+
+        x = torch.nn.functional.pad(x.unsqueeze(1), (1, 0), "reflect")
+        return torch.nn.functional.conv1d(x, self.filter).squeeze(1)
+
+
+
+class ResNetSpeakerEncoder(nn.Module):
+    """This is copied from 🐸TTS to remove it from the dependencies.
+    """
+
+    # pylint: disable=W0102
+    def __init__(
+        self,
+        input_dim=64,
+        proj_dim=512,
+        layers=[3, 4, 6, 3],
+        num_filters=[32, 64, 128, 256],
+        encoder_type="ASP",
+        log_input=False,
+        use_torch_spec=False,
+        audio_config=None,
+    ):
+        super(ResNetSpeakerEncoder, self).__init__()
+
+        self.encoder_type = encoder_type
+        self.input_dim = input_dim
+        self.log_input = log_input
+        self.use_torch_spec = use_torch_spec
+        self.audio_config = audio_config
+        self.proj_dim = proj_dim
+
+        self.conv1 = nn.Conv2d(1, num_filters[0], kernel_size=3, stride=1, padding=1)
+        self.relu = nn.ReLU(inplace=True)
+        self.bn1 = nn.BatchNorm2d(num_filters[0])
+
+        self.inplanes = num_filters[0]
+        self.layer1 = self.create_layer(SEBasicBlock, num_filters[0], layers[0])
+        self.layer2 = self.create_layer(SEBasicBlock, num_filters[1], layers[1], stride=(2, 2))
+        self.layer3 = self.create_layer(SEBasicBlock, num_filters[2], layers[2], stride=(2, 2))
+        self.layer4 = self.create_layer(SEBasicBlock, num_filters[3], layers[3], stride=(2, 2))
+
+        self.instancenorm = nn.InstanceNorm1d(input_dim)
+
+        if self.use_torch_spec:
+            self.torch_spec = torch.nn.Sequential(
+                PreEmphasis(audio_config["preemphasis"]),
+                torchaudio.transforms.MelSpectrogram(
+                    sample_rate=audio_config["sample_rate"],
+                    n_fft=audio_config["fft_size"],
+                    win_length=audio_config["win_length"],
+                    hop_length=audio_config["hop_length"],
+                    window_fn=torch.hamming_window,
+                    n_mels=audio_config["num_mels"],
+                ),
+            )
+
+        else:
+            self.torch_spec = None
+
+        outmap_size = int(self.input_dim / 8)
+
+        self.attention = nn.Sequential(
+            nn.Conv1d(num_filters[3] * outmap_size, 128, kernel_size=1),
+            nn.ReLU(),
+            nn.BatchNorm1d(128),
+            nn.Conv1d(128, num_filters[3] * outmap_size, kernel_size=1),
+            nn.Softmax(dim=2),
+        )
+
+        if self.encoder_type == "SAP":
+            out_dim = num_filters[3] * outmap_size
+        elif self.encoder_type == "ASP":
+            out_dim = num_filters[3] * outmap_size * 2
+        else:
+            raise ValueError("Undefined encoder")
+
+        self.fc = nn.Linear(out_dim, proj_dim)
+
+        self._init_layers()
+
+    def _init_layers(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+    def create_layer(self, block, planes, blocks, stride=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, downsample))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(block(self.inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    # pylint: disable=R0201
+    def new_parameter(self, *size):
+        out = nn.Parameter(torch.FloatTensor(*size))
+        nn.init.xavier_normal_(out)
+        return out
+
+    def forward(self, x, l2_norm=False):
+        """Forward pass of the model.
+
+        Args:
+            x (Tensor): Raw waveform signal or spectrogram frames. If input is a waveform, `torch_spec` must be `True`
+                to compute the spectrogram on-the-fly.
+            l2_norm (bool): Whether to L2-normalize the outputs.
+
+        Shapes:
+            - x: :math:`(N, 1, T_{in})` or :math:`(N, D_{spec}, T_{in})`
+        """
+        x.squeeze_(1)
+        # if you torch spec compute it otherwise use the mel spec computed by the AP
+        if self.use_torch_spec:
+            x = self.torch_spec(x)
+
+        if self.log_input:
+            x = (x + 1e-6).log()
+        x = self.instancenorm(x).unsqueeze(1)
+
+        x = self.conv1(x)
+        x = self.relu(x)
+        x = self.bn1(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = x.reshape(x.size()[0], -1, x.size()[-1])
+
+        w = self.attention(x)
+
+        if self.encoder_type == "SAP":
+            x = torch.sum(x * w, dim=2)
+        elif self.encoder_type == "ASP":
+            mu = torch.sum(x * w, dim=2)
+            sg = torch.sqrt((torch.sum((x**2) * w, dim=2) - mu**2).clamp(min=1e-5))
+            x = torch.cat((mu, sg), 1)
+
+        x = x.view(x.size()[0], -1)
+        x = self.fc(x)
+
+        if l2_norm:
+            x = torch.nn.functional.normalize(x, p=2, dim=1)
+        return x
+
+    def load_checkpoint(
+        self,
+        checkpoint_path: str,
+        eval: bool = False,
+        use_cuda: bool = False,
+        criterion=None,
+        cache=False,
+    ):
+        state = load_fsspec(checkpoint_path, map_location=torch.device("cpu"), cache=cache)
+        try:
+            self.load_state_dict(state["model"])
+            print(" > Model fully restored. ")
+        except (KeyError, RuntimeError) as error:
+            # If eval raise the error
+            if eval:
+                raise error
+
+            print(" > Partial model initialization.")
+            model_dict = self.state_dict()
+            model_dict = set_init_dict(model_dict, state["model"])
+            self.load_state_dict(model_dict)
+            del model_dict
+
+        # load the criterion for restore_path
+        if criterion is not None and "criterion" in state:
+            try:
+                criterion.load_state_dict(state["criterion"])
+            except (KeyError, RuntimeError) as error:
+                print(" > Criterion load ignored because of:", error)
+
+        if use_cuda:
+            self.cuda()
+            if criterion is not None:
+                criterion = criterion.cuda()
+
+        if eval:
+            self.eval()
+            assert not self.training
+
+        if not eval:
+            return criterion, state["step"]
+        return criterion
+
+class HifiDecoder(torch.nn.Module):
+    def __init__(
+        self,
+        input_sample_rate=22050,
+        output_sample_rate=24000,
+        output_hop_length=256,
+        ar_mel_length_compression=1024,
+        decoder_input_dim=1024,
+        resblock_type_decoder="1",
+        resblock_dilation_sizes_decoder=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+        resblock_kernel_sizes_decoder=[3, 7, 11],
+        upsample_rates_decoder=[8, 8, 2, 2],
+        upsample_initial_channel_decoder=512,
+        upsample_kernel_sizes_decoder=[16, 16, 4, 4],
+        d_vector_dim=512,
+        cond_d_vector_in_each_upsampling_layer=True,
+        speaker_encoder_audio_config={
+            "fft_size": 512,
+            "win_length": 400,
+            "hop_length": 160,
+            "sample_rate": 16000,
+            "preemphasis": 0.97,
+            "num_mels": 64,
+        },
+    ):
+        super().__init__()
+        self.input_sample_rate = input_sample_rate
+        self.output_sample_rate = output_sample_rate
+        self.output_hop_length = output_hop_length
+        self.ar_mel_length_compression = ar_mel_length_compression
+        self.speaker_encoder_audio_config = speaker_encoder_audio_config
+        self.waveform_decoder = HifiganGenerator(
+            decoder_input_dim,
+            1,
+            resblock_type_decoder,
+            resblock_dilation_sizes_decoder,
+            resblock_kernel_sizes_decoder,
+            upsample_kernel_sizes_decoder,
+            upsample_initial_channel_decoder,
+            upsample_rates_decoder,
+            inference_padding=0,
+            cond_channels=d_vector_dim,
+            conv_pre_weight_norm=False,
+            conv_post_weight_norm=False,
+            conv_post_bias=False,
+            cond_in_each_up_layer=cond_d_vector_in_each_upsampling_layer,
+        )
+        self.speaker_encoder = ResNetSpeakerEncoder(
+            input_dim=64,
+            proj_dim=512,
+            log_input=True,
+            use_torch_spec=True,
+            audio_config=speaker_encoder_audio_config,
+        )
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    def forward(self, latents, g=None):
+        """
+        Args:
+            x (Tensor): feature input tensor (GPT latent).
+            g (Tensor): global conditioning input tensor.
+
+        Returns:
+            Tensor: output waveform.
+
+        Shapes:
+            x: [B, C, T]
+            Tensor: [B, 1, T]
+        """
+
+        z = torch.nn.functional.interpolate(
+            latents.transpose(1, 2),
+            scale_factor=[self.ar_mel_length_compression / self.output_hop_length],
+            mode="linear",
+        ).squeeze(1)
+        # upsample to the right sr
+        if self.output_sample_rate != self.input_sample_rate:
+            z = torch.nn.functional.interpolate(
+                z,
+                scale_factor=[self.output_sample_rate / self.input_sample_rate],
+                mode="linear",
+            ).squeeze(0)
+        o = self.waveform_decoder(z, g=g)
+        return o
+
+    @torch.no_grad()
+    def inference(self, c, g):
+        """
+        Args:
+            x (Tensor): feature input tensor (GPT latent).
+            g (Tensor): global conditioning input tensor.
+
+        Returns:
+            Tensor: output waveform.
+
+        Shapes:
+            x: [B, C, T]
+            Tensor: [B, 1, T]
+        """
+        return self.forward(c, g=g)
+
+    def load_checkpoint(
+        self, checkpoint_path, eval=False
+    ):  # pylint: disable=unused-argument, redefined-builtin
+        state = load_fsspec(checkpoint_path, map_location=torch.device("cpu"))
+        # remove unused keys
+        state = state["model"]
+        states_keys = list(state.keys())
+        for key in states_keys:
+            if "waveform_decoder." not in key and "speaker_encoder." not in key:
+                del state[key]
+
+        self.load_state_dict(state)
+        if eval:
+            self.eval()
+            assert not self.training
+            self.waveform_decoder.remove_weight_norm()

TTS/tts/models/xtts.py

@@ -13,9 +13,12 @@ from TTS.tts.layers.xtts.diffusion import SpacedDiffusion, get_named_beta_schedu
 from TTS.tts.layers.xtts.gpt import GPT
 from TTS.tts.layers.xtts.tokenizer import VoiceBpeTokenizer
 from TTS.tts.layers.xtts.vocoder import UnivNetGenerator
+from TTS.tts.layers.xtts.hifigan_decoder import HifiDecoder
+from TTS.tts.layers.xtts.stream_generator import init_stream_support
 from TTS.tts.models.base_tts import BaseTTS
 from TTS.utils.io import load_fsspec
 
+init_stream_support()
 
 def load_audio(audiopath, sr=22050):
     """
@@ -266,6 +269,15 @@ class XttsArgs(Coqpit):
     diff_layer_drop: int = 0
     diff_unconditioned_percentage: int = 0
 
+    # HifiGAN Decoder params
+    input_sample_rate: int = 22050
+    output_sample_rate: int = 24000
+    output_hop_length: int = 256
+    ar_mel_length_compression: int = 1024
+    decoder_input_dim: int = 1024
+    d_vector_dim: int = 512
+    cond_d_vector_in_each_upsampling_layer: bool = True
+
     # constants
     duration_const: int = 102400
 
@@ -322,6 +334,16 @@ class Xtts(BaseTTS):
                 stop_audio_token=self.args.gpt_stop_audio_token,
             )
 
+        self.hifigan_decoder = HifiDecoder(
+            input_sample_rate=self.args.input_sample_rate,
+            output_sample_rate=self.args.output_sample_rate,
+            output_hop_length=self.args.output_hop_length,
+            ar_mel_length_compression=self.args.ar_mel_length_compression,
+            decoder_input_dim=self.args.decoder_input_dim,
+            d_vector_dim=self.args.d_vector_dim,
+            cond_d_vector_in_each_upsampling_layer=self.args.cond_d_vector_in_each_upsampling_layer,
+        )
+
         self.diffusion_decoder = DiffusionTts(
             model_channels=self.args.diff_model_channels,
             num_layers=self.args.diff_num_layers,
@@ -393,16 +415,30 @@ class Xtts(BaseTTS):
             diffusion_latent = diffusion.get_conditioning(diffusion_conds)
         return diffusion_latent
 
+    def get_speaker_embedding(
+        self,
+        audio_path
+    ):
+        wav, sr = torchaudio.load(audio_path)
+        spk_waveform = torchaudio.functional.resample(
+            wav,
+            22050,
+            self.hifigan_decoder.speaker_encoder_audio_config["sample_rate"],
+        ).to(self.device)
+        speaker_embedding = self.hifigan_decoder.speaker_encoder.forward(
+            spk_waveform, l2_norm=True
+        ).unsqueeze(-1).to(self.device)
+        return speaker_embedding
+
     def get_conditioning_latents(
         self,
         audio_path,
         gpt_cond_len=3,
     ):
         gpt_cond_latents = self.get_gpt_cond_latents(audio_path, length=gpt_cond_len)  # [1, 1024, T]
-        diffusion_cond_latents = self.get_diffusion_cond_latents(
+        diffusion_cond_latents = self.get_diffusion_cond_latents(audio_path)
-            audio_path,
+        speaker_embedding = self.get_speaker_embedding(audio_path)
-        )
+        return gpt_cond_latents.to(self.device), diffusion_cond_latents.to(self.device), speaker_embedding
-        return gpt_cond_latents.to(self.device), diffusion_cond_latents.to(self.device)
 
     def synthesize(self, text, config, speaker_wav, language, **kwargs):
         """Synthesize speech with the given input text.
@@ -469,6 +505,7 @@ class Xtts(BaseTTS):
         cond_free_k=2,
         diffusion_temperature=1.0,
         decoder_sampler="ddim",
+        use_hifigan=True,
         **hf_generate_kwargs,
     ):
         """
@@ -535,8 +572,10 @@ class Xtts(BaseTTS):
         (
             gpt_cond_latent,
             diffusion_conditioning,
+            speaker_embedding
         ) = self.get_conditioning_latents(audio_path=ref_audio_path, gpt_cond_len=gpt_cond_len)
 
+        if not use_hifigan:
             diffuser = load_discrete_vocoder_diffuser(
                 desired_diffusion_steps=decoder_iterations,
                 cond_free=cond_free,
@@ -561,8 +600,6 @@ class Xtts(BaseTTS):
                     output_attentions=False,
                     **hf_generate_kwargs,
                 )
-
-            with self.lazy_load_model(self.gpt) as gpt:
                 expected_output_len = torch.tensor(
                     [gpt_codes.shape[-1] * self.gpt.code_stride_len], device=text_tokens.device
                 )
@@ -587,6 +624,10 @@ class Xtts(BaseTTS):
                         gpt_latents = gpt_latents[:, :k]
                         break
             
+            if use_hifigan:
+                with self.lazy_load_model(self.hifigan_decoder) as hifigan_decoder:
+                    wav = hifigan_decoder(gpt_latents, g=speaker_embedding)
+            else:
                 with self.lazy_load_model(self.diffusion_decoder) as diffusion:
                     mel = do_spectrogram_diffusion(
                         diffusion,
@@ -600,6 +641,102 @@ class Xtts(BaseTTS):
 
         return {"wav": wav.cpu().numpy().squeeze()}
 
+    def inference_speaker_cond(self, ref_audio_path, gpt_cond_len=3):
+        (
+            gpt_cond_latent,
+            diffusion_conditioning,
+            speaker_embedding
+        ) = self.get_conditioning_latents(audio_path=ref_audio_path, gpt_cond_len=3)
+        return {
+            "gpt_cond_latent": gpt_cond_latent,
+            "speaker_embedding": speaker_embedding,
+            "diffusion_conditioning": diffusion_conditioning,
+        }
+
+    def handle_chunks(self, wav_gen, wav_gen_prev, wav_overlap, overlap_len):
+        """Handle chunk formatting in streaming mode"""
+        wav_chunk = wav_gen[:-overlap_len]
+        if wav_gen_prev is not None:
+            wav_chunk = wav_gen[(wav_gen_prev.shape[0] - overlap_len) : -overlap_len]
+        if wav_overlap is not None:
+            crossfade_wav = wav_chunk[:overlap_len]
+            crossfade_wav = crossfade_wav * torch.linspace(0.0, 1.0, overlap_len).to(crossfade_wav.device)
+            wav_chunk[:overlap_len] = wav_overlap * torch.linspace(1.0, 0.0, overlap_len).to(wav_overlap.device)
+            wav_chunk[:overlap_len] += crossfade_wav
+        wav_overlap = wav_gen[-overlap_len:]
+        wav_gen_prev = wav_gen
+        return wav_chunk, wav_gen_prev, wav_overlap
+
+    def inference_stream(
+        self,
+        text,
+        language,
+        gpt_cond_latent,
+        speaker_embedding,
+        diffusion_conditioning,
+        # Streaming
+        stream_chunk_size=20,
+        overlap_wav_len=1024,
+        # GPT inference
+        temperature=0.65,
+        length_penalty=1,
+        repetition_penalty=2.0,
+        top_k=50,
+        top_p=0.85,
+        gpt_cond_len=4,
+        do_sample=True,
+        # Decoder inference
+        decoder_iterations=100,
+        cond_free=True,
+        cond_free_k=2,
+        diffusion_temperature=1.0,
+        decoder_sampler="ddim",
+        **hf_generate_kwargs,
+    ):
+        text = f"[{language}]{text.strip().lower()}"
+        text_tokens = torch.IntTensor(self.tokenizer.encode(text, lang=language)).unsqueeze(0).to(self.device)
+
+        fake_inputs = self.gpt.compute_embeddings(
+            gpt_cond_latent.to(self.device),
+            text_tokens,
+        )
+        gpt_generator = self.gpt.get_generator(
+            fake_inputs=fake_inputs,
+            top_k=top_k,
+            top_p=top_p,
+            temperature=temperature,
+            do_sample=do_sample,
+            num_beams=1,
+            num_return_sequences=1,
+            length_penalty=float(length_penalty),
+            repetition_penalty=float(repetition_penalty),
+            output_attentions=False,
+            output_hidden_states=True,
+        )
+
+        last_tokens = []
+        all_latents = []
+        wav_gen_prev = None
+        wav_overlap = None
+        is_end = False
+
+        while not is_end:
+            try:
+                x, latent = next(gpt_generator)
+                last_tokens += [x]
+                all_latents += [latent]
+            except StopIteration:
+                is_end = True
+
+            if is_end or (stream_chunk_size > 0 and len(last_tokens) >= stream_chunk_size):
+                gpt_latents = torch.cat(all_latents, dim=0)[None, :]
+                wav_gen = self.hifigan_decoder(gpt_latents, g=speaker_embedding.to(self.device))
+                wav_chunk, wav_gen_prev, wav_overlap = self.handle_chunks(
+                    wav_gen.squeeze(), wav_gen_prev, wav_overlap, overlap_wav_len
+                )
+                last_tokens = []
+                yield wav_chunk
+
     def forward(self):
         raise NotImplementedError("XTTS Training is not implemented")
 
@@ -616,7 +753,14 @@ class Xtts(BaseTTS):
         super().eval()
 
     def load_checkpoint(
-        self, config, checkpoint_dir=None, checkpoint_path=None, vocab_path=None, eval=False, strict=True
+        self,
+        config,
+        checkpoint_dir=None,
+        checkpoint_path=None, 
+        vocab_path=None,
+        eval=False,
+        strict=True,
+        use_deepspeed=False
     ):
         """
         Loads a checkpoint from disk and initializes the model's state and tokenizer.
@@ -636,8 +780,8 @@ class Xtts(BaseTTS):
         model_path = checkpoint_path or os.path.join(checkpoint_dir, "model.pth")
         vocab_path = vocab_path or os.path.join(checkpoint_dir, "vocab.json")
 
-        if os.path.exists(os.path.join(checkpoint_dir, "vocab.json")):
+        if os.path.exists(vocab_path):
-            self.tokenizer = VoiceBpeTokenizer(vocab_file=os.path.join(checkpoint_dir, "vocab.json"))
+            self.tokenizer = VoiceBpeTokenizer(vocab_file=vocab_path)
 
         self.init_models()
         if eval:
@@ -645,10 +789,11 @@ class Xtts(BaseTTS):
         self.load_state_dict(load_fsspec(model_path, map_location=self.device)["model"], strict=strict)
 
         if eval:
-            self.gpt.init_gpt_for_inference(kv_cache=self.args.kv_cache)
+            self.gpt.init_gpt_for_inference(kv_cache=self.args.kv_cache, use_deepspeed=use_deepspeed)
             self.gpt.eval()
             self.diffusion_decoder.eval()
             self.vocoder.eval()
+            self.hifigan_decoder.eval()
 
     def train_step(self):
         raise NotImplementedError("XTTS Training is not implemented")