Add pitch and prosody encoder suport for the conditional module

TTS/tts/configs/vits_config.py

@@ -118,8 +118,8 @@ class VitsConfig(BaseTTSConfig):
     speaker_classifier_loss_alpha: float = 2.0
     emotion_classifier_loss_alpha: float = 4.0
     prosody_encoder_kl_loss_alpha: float = 5.0
-    disc_latent_loss_alpha: float = 5.0
-    gen_latent_loss_alpha: float = 5.0
+    disc_latent_loss_alpha: float = 1.0
+    gen_latent_loss_alpha: float = 1.0
     feat_latent_loss_alpha: float = 108.0
     pitch_loss_alpha: float = 5.0
     z_decoder_loss_alpha: float = 45.0

TTS/tts/models/vits.py

@@ -788,7 +788,7 @@ class Vits(BaseTTS):
             self.args.dropout_p_text_encoder,
             language_emb_dim=self.embedded_language_dim,
             emotion_emb_dim=self.args.emotion_embedding_dim if not self.args.use_noise_scale_predictor else 0,
-            prosody_emb_dim=self.args.prosody_embedding_dim if not self.args.use_noise_scale_predictor and not self.args.use_encoder_conditional_module else 0,
+            prosody_emb_dim=self.args.prosody_embedding_dim if not self.args.use_noise_scale_predictor and not self.args.use_encoder_conditional_module and not self.args.use_z_decoder else 0,
             pitch_dim=self.args.pitch_embedding_dim if self.args.use_pitch and self.args.use_pitch_on_enc_input else 0,
         )
 
@@ -827,7 +827,7 @@ class Vits(BaseTTS):
         ) and not self.args.use_noise_scale_predictor:
             dp_extra_inp_dim += self.args.emotion_embedding_dim
 
-        if self.args.use_prosody_encoder and not self.args.use_noise_scale_predictor and not self.args.use_encoder_conditional_module:
+        if self.args.use_prosody_encoder and not self.args.use_noise_scale_predictor and not self.args.use_encoder_conditional_module and not self.args.use_z_decoder:
             dp_extra_inp_dim += self.args.prosody_embedding_dim
 
         if self.args.use_pitch and self.args.use_pitch_on_enc_input:
@@ -896,7 +896,7 @@ class Vits(BaseTTS):
                 self.args.pitch_predictor_hidden_channels,
                 self.args.pitch_predictor_kernel_size,
                 self.args.pitch_predictor_dropout_p,
-                cond_channels=dp_cond_embedding_dim,
+                cond_channels=self.cond_embedding_dim,
                 language_emb_dim=self.embedded_language_dim,
             )
 
@@ -1477,7 +1477,7 @@ class Vits(BaseTTS):
         if self.args.use_pitch and self.args.use_pitch_on_enc_input:
             if alignments is None:
                 raise RuntimeError(" [!] For condition the pitch on the Text Encoder you need to provide external alignments !")
-            pitch_loss, gt_avg_pitch_emb, _ = self.forward_pitch_predictor(x, x_lengths, pitch, alignments.sum(3), g_dp)
+            pitch_loss, gt_avg_pitch_emb, _ = self.forward_pitch_predictor(x, x_lengths, pitch, alignments.sum(3), g)
 
         # posterior encoder
         z, m_q, logs_q, y_mask = self.posterior_encoder(y, y_lengths, g=g)
@@ -1514,7 +1514,7 @@ class Vits(BaseTTS):
             x_lengths,
             lang_emb=lang_emb,
             emo_emb=eg if not self.args.use_noise_scale_predictor else None,
-            pros_emb=pros_emb if not self.args.use_noise_scale_predictor and not self.args.use_encoder_conditional_module else None,
+            pros_emb=pros_emb if not self.args.use_noise_scale_predictor and not self.args.use_encoder_conditional_module and not self.args.use_z_decoder else None,
             pitch_emb=gt_avg_pitch_emb if self.args.use_pitch and self.args.use_pitch_on_enc_input else None,
         )
 
@@ -1536,10 +1536,20 @@ class Vits(BaseTTS):
 
         outputs, attn = self.forward_mas(outputs, z_p, m_p, logs_p, x, x_mask, y_mask, g=g_dp, lang_emb=lang_emb)
 
+        z_p_avg = None
+        if self.args.use_latent_discriminator:
+            # average the z_p for the latent discriminator
+            z_p_avg = average_over_durations(z_p, attn.sum(3).squeeze())
+
         conditional_module_loss = None
+        new_m_p = None
         if self.args.use_encoder_conditional_module:
             g_cond = None
             cond_module_input = x
+            if self.args.use_pitch and not self.args.use_pitch_on_enc_input:
+                pitch_loss, gt_avg_pitch_emb, _ = self.forward_pitch_predictor(cond_module_input, x_lengths, pitch, attn.sum(3), g)
+                cond_module_input = cond_module_input + gt_avg_pitch_emb
+
             if self.args.use_prosody_encoder:
                 if g_cond is None:
                     g_cond = pros_emb
@@ -1549,18 +1559,17 @@ class Vits(BaseTTS):
             if g_cond is not None:
                 cond_module_input = torch.cat((cond_module_input, g_cond.expand(-1, -1, cond_module_input.size(2))), dim=1)
 
-            new_m_p = self.encoder_conditional_module(cond_module_input, x_mask)
-            z_p_avg = average_over_durations(z_p, attn.sum(3).squeeze()).detach()
-            conditional_module_loss = torch.nn.functional.l1_loss(new_m_p * x_mask, z_p_avg)
-            
-        if self.args.use_pitch and not self.args.use_pitch_on_enc_input:
-            pitch_loss, gt_avg_pitch_emb, _ = self.forward_pitch_predictor(m_p, x_lengths, pitch, attn.sum(3), g_dp)
-            m_p = m_p + gt_avg_pitch_emb
+            new_m_p = self.encoder_conditional_module(cond_module_input, x_mask) * x_mask
+            if z_p_avg is None:
+                z_p_avg = average_over_durations(z_p, attn.sum(3).squeeze()).detach()
+            else:
+                z_p_avg = z_p_avg.detach()
 
-        z_p_avg = None
-        if self.args.use_latent_discriminator:
-            # average the z_p for the latent discriminator
-            z_p_avg = average_over_durations(z_p, attn.sum(3).squeeze())
+            conditional_module_loss = torch.nn.functional.l1_loss(new_m_p, z_p_avg)
+
+        if self.args.use_pitch and not self.args.use_pitch_on_enc_input and not self.args.use_z_decoder and not self.args.use_encoder_conditional_module:
+            pitch_loss, gt_avg_pitch_emb, _ = self.forward_pitch_predictor(m_p, x_lengths, pitch, attn.sum(3), g)
+            m_p = m_p + gt_avg_pitch_emb
 
         # expand prior
         m_p_expanded = torch.einsum("klmn, kjm -> kjn", [attn, m_p])
@@ -1568,7 +1577,12 @@ class Vits(BaseTTS):
 
         z_decoder_loss = None
         if self.args.use_z_decoder:
-            x_expanded = torch.einsum("klmn, kjm -> kjn", [attn, x])
+            cond_x = x
+            if self.args.use_pitch and not self.args.use_pitch_on_enc_input:
+                pitch_loss, gt_avg_pitch_emb, _ = self.forward_pitch_predictor(cond_x, x_lengths, pitch, attn.sum(3), g)
+                cond_x = cond_x + gt_avg_pitch_emb
+
+            x_expanded = torch.einsum("klmn, kjm -> kjn", [attn, cond_x])
             # prepare the conditional emb
             g_dec = g
             if eg is not None and (self.args.use_emotion_embedding or self.args.use_external_emotions_embeddings):
@@ -1646,7 +1660,7 @@ class Vits(BaseTTS):
             {
                 "model_outputs": o,
                 "alignments": attn.squeeze(1),
-                "m_p_unexpanded": m_p,
+                "m_p_unexpanded": m_p if new_m_p is None else new_m_p,
                 "z_p_avg": z_p_avg,
                 "m_p": m_p_expanded,
                 "logs_p": logs_p_expanded,
@@ -1774,14 +1788,14 @@ class Vits(BaseTTS):
 
         pred_avg_pitch_emb = None
         if self.args.use_pitch and self.args.use_pitch_on_enc_input:
-            _, _, pred_avg_pitch_emb = self.forward_pitch_predictor(x, x_lengths, g_pp=g_dp, pitch_transform=pitch_transform)
+            _, _, pred_avg_pitch_emb = self.forward_pitch_predictor(x, x_lengths, g_pp=g, pitch_transform=pitch_transform)
 
         x, m_p, logs_p, x_mask = self.text_encoder(
             x,
             x_lengths,
             lang_emb=lang_emb,
             emo_emb=eg if not self.args.use_noise_scale_predictor else None,
-            pros_emb=pros_emb if not self.args.use_noise_scale_predictor and not self.args.use_encoder_conditional_module else None,
+            pros_emb=pros_emb if not self.args.use_noise_scale_predictor and not self.args.use_encoder_conditional_module and not self.args.use_z_decoder else None,
             pitch_emb=pred_avg_pitch_emb if self.args.use_pitch and self.args.use_pitch_on_enc_input else None,
         )
 
@@ -1811,19 +1825,22 @@ class Vits(BaseTTS):
         attn_mask = x_mask * y_mask.transpose(1, 2)  # [B, 1, T_enc] * [B, T_dec, 1]
         attn = generate_path(w_ceil.squeeze(1), attn_mask.squeeze(1).transpose(1, 2))
 
-        if self.args.use_pitch and not self.args.use_pitch_on_enc_input:
-            _, _, pred_avg_pitch_emb = self.forward_pitch_predictor(m_p, x_lengths, g_pp=g_dp, pitch_transform=pitch_transform)
+        if self.args.use_pitch and not self.args.use_pitch_on_enc_input and not self.args.use_z_decoder and not self.args.use_encoder_conditional_module:
+            _, _, pred_avg_pitch_emb = self.forward_pitch_predictor(m_p, x_lengths, g_pp=g, pitch_transform=pitch_transform)
             m_p = m_p + pred_avg_pitch_emb
 
         if self.args.use_encoder_conditional_module:
             g_cond = None
             cond_module_input = x
+            if self.args.use_pitch and not self.args.use_pitch_on_enc_input:
+                _, _, pred_avg_pitch_emb = self.forward_pitch_predictor(cond_module_input, x_lengths, g_pp=g, pitch_transform=pitch_transform)
+                cond_module_input = cond_module_input + pred_avg_pitch_emb
+
             if self.args.use_prosody_encoder:
                 if g_cond is None:
                     g_cond = pros_emb
                 else:
                     g_cond = torch.cat([g_cond, pros_emb], dim=1)  # [b, h1+h2, 1]
-
             if g_cond is not None:
                 cond_module_input = torch.cat((cond_module_input, g_cond.expand(-1, -1, cond_module_input.size(2))), dim=1)
             m_p = self.encoder_conditional_module(cond_module_input, x_mask)
@@ -1850,14 +1867,20 @@ class Vits(BaseTTS):
             z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * self.inference_noise_scale
 
         if self.args.use_z_decoder:
-            x_expanded = torch.matmul(attn.transpose(1, 2), x.transpose(1, 2)).transpose(1, 2)
+            cond_x = x
+            if self.args.use_pitch and not self.args.use_pitch_on_enc_input:
+                _, _, pred_avg_pitch_emb = self.forward_pitch_predictor(cond_x, x_lengths, g_pp=g, pitch_transform=pitch_transform)
+                cond_x = cond_x + pred_avg_pitch_emb
+
+            x_expanded = torch.matmul(attn.transpose(1, 2), cond_x.transpose(1, 2)).transpose(1, 2)
             # prepare the conditional emb
             g_dec = g
             if eg is not None and (self.args.use_emotion_embedding or self.args.use_external_emotions_embeddings):
                 if g_dec is None:
                     g_dec = eg
                 else:
-                    g_dec = torch.cat([g_dec, eg], dim=1)  # [b, h1+h2, 1]
+                    g_dec = torch.cat([g_dec, eg], dim=1)  # [b, h1+h2, 1]+
+    
             if self.args.use_prosody_encoder:
                 if g_dec is None:
                     g_dec = pros_emb
@@ -2653,4 +2676,4 @@ class VitsCharacters(BaseCharacters):
             blank=self._blank,
             is_unique=False,
             is_sorted=True,
-        )
+        )

tests/tts_tests/test_vits_speaker_emb_with_pitch_predictor.py

@@ -48,12 +48,15 @@ config.model_args.alignments_cache_path = "tests/data/ljspeech/mas_alignments/al
 
 # pitch predictor
 config.model_args.use_pitch = True
-config.model_args.use_pitch_on_enc_input = True
+config.model_args.use_pitch_on_enc_input = False
 config.model_args.pitch_embedding_dim = 2
-config.model_args.condition_dp_on_speaker = True
+config.model_args.condition_dp_on_speaker = False
 
 
-config.model_args.use_latent_discriminator = True
+config.model_args.use_encoder_conditional_module = True
+config.model_args.use_z_decoder = False
+
+config.model_args.use_latent_discriminator = False
 
 config.save_json(config_path)
 # train the model for one epoch

tests/tts_tests/test_vits_speaker_emb_with_prosody_encoder.py

@@ -46,7 +46,8 @@ config.model_args.d_vector_dim = 128
 config.model_args.use_prosody_encoder = True
 config.model_args.prosody_embedding_dim = 64
 
-config.model_args.use_encoder_conditional_module = True
+config.model_args.use_z_decoder = True
+config.model_args.use_encoder_conditional_module = False
 
 # active classifier
 config.model_args.external_emotions_embs_file = "tests/data/ljspeech/speakers.json"