update svc inference and webui

example/infer_svc.sh

@@ -6,7 +6,7 @@ root_dir=$(dirname "$script_dir")
 cd $root_dir || exit
 export PYTHONPATH=$root_dir:$PYTHONPATH
 
-model_path=pretrained_models/SoulX-Singer-SVC/model.pt
+model_path=pretrained_models/SoulX-Singer/model-svc.pt
 config=soulxsinger/config/soulxsinger.yaml
 prompt_wav_path=example/audio/zh_prompt.mp3
 target_wav_path=example/audio/music.mp3

soulxsinger/models/soulxsinger_svc.py

@@ -212,6 +212,20 @@ class SoulXSingerSVC(nn.Module):
         else:
             pitch_shift = pitch_shift
 
+        # if target audio is less than 30 seconds, infer the whole audio
+        if gt_wav.shape[-1] < 30 * self.audio_cfg.sample_rate:
+            generated_audio = self.infer_segment(
+                pt_wav=pt_wav,
+                gt_wav=gt_wav,
+                pt_f0=pt_f0,
+                gt_f0=gt_f0,
+                pitch_shift=pitch_shift,
+                n_steps=n_steps,
+                cfg=cfg,
+            )
+            return generated_audio, pitch_shift
+
+        # if target audio is longer than 30 seconds, build vocal segments and infer each segment
         generated_audio = []
 
         f0_rate = self.audio_cfg.sample_rate // self.audio_cfg.hop_size
@@ -264,7 +278,8 @@ class SoulXSingerSVC(nn.Module):
 
     def infer_segment(self, pt_wav, gt_wav, pt_f0, gt_f0, pitch_shift=0, n_steps=32, cfg=3):
         pt_mel = self.mel(pt_wav)
-        len_prompt_mel = pt_f0.shape[1]
+        len_prompt_mel = pt_mel.shape[1]
+        pt_f0 = F.pad(pt_f0, (0, 0, 0, max(0, len_prompt_mel - pt_f0.shape[1])))[:, :len_prompt_mel]
 
         f0_course_pt = self.f0_to_coarse(pt_f0)
         f0_course_gt = self.f0_to_coarse(gt_f0, f0_shift=pitch_shift * 5)
@@ -272,12 +287,13 @@ class SoulXSingerSVC(nn.Module):
 
         pt_content_feat = self.whisper_encoder.encode(pt_wav, sr=self.audio_cfg.sample_rate)
         gt_content_feat = self.whisper_encoder.encode(gt_wav, sr=self.audio_cfg.sample_rate)
+        t_pt, t_gt = f0_course_pt.shape[1], f0_course_gt.shape[1]
+        pt_content_feat = F.pad(pt_content_feat, (0, 0, 0, max(0, t_pt - pt_content_feat.shape[1])))[:, :t_pt, :]
+        gt_content_feat = F.pad(gt_content_feat, (0, 0, 0, max(0, t_gt - gt_content_feat.shape[1])))[:, :t_gt, :]
+
         content_feat = torch.cat([pt_content_feat, gt_content_feat], 1)
 
         f0_feat = self.f0_encoder(f0_course)
-        min_len = min(content_feat.shape[1], f0_feat.shape[1])
-        content_feat = content_feat[:, :min_len, :]
-        f0_feat = f0_feat[:, :min_len, :]
         features = content_feat + f0_feat
         
         gt_decoder_inp = features[:, len_prompt_mel:, :]

webui_svc.py

@@ -60,7 +60,7 @@ _I18N = dict(
         zh="上传 Prompt 与 Target 音频，并配置相关参数",
     ),
     instruction_p2=dict(
-        en="Click the button to start singing voice conversion.",
+        en="Click「🎤Singing Voice Conversion」to start singing voice conversion.",
         zh="点击「🎤歌声转换」开始最终生成。",
     ),
 	tips_title=dict(en="Tips", zh="提示"),
@@ -145,7 +145,7 @@ class AppState:
 
 		self.svc_config = load_config("soulxsinger/config/soulxsinger.yaml")
 		self.svc_model = build_svc_model(
-			model_path="pretrained_models/SoulX-Singer-SVC/model.pt",
+			model_path="pretrained_models/SoulX-Singer/model-svc.pt",
 			config=self.svc_config,
 			device=self.device,
 		)
@@ -194,8 +194,6 @@ class AppState:
 
 			args = Args()
 			args.device = self.device
-			args.model_path = "soulx-singer-svc.pt"
-			args.config = "soulxsinger/config/soulxsinger.yaml"
 			args.prompt_wav_path = str(prompt_wav_path)
 			args.target_wav_path = str(target_wav_path)
 			args.prompt_f0_path = str(prompt_f0_path)
@@ -305,7 +303,30 @@ def _start_svc(prompt_audio, target_audio, prompt_vocal_sep, target_vocal_sep, a
 
 
 def render_interface() -> gr.Blocks:
-	with gr.Blocks(title="SoulX-Singer SVC Demo", theme=gr.themes.Default()) as page:
+	with gr.Blocks(title="SoulX-Singer-SVC Demo", theme=gr.themes.Default()) as page:
+		gr.HTML(
+            '<div style="'
+            'text-align: center; '
+            'padding: 1.25rem 0 1.5rem; '
+            'margin-bottom: 0.5rem;'
+            '">'
+            '<div style="'
+            'display: inline-block; '
+            'font-size: 1.75rem; '
+            'font-weight: 700; '
+            'letter-spacing: 0.02em; '
+            'color: #1a1a2e; '
+            'line-height: 1.3;'
+            '">SoulX-Singer-SVC</div>'
+            '<div style="'
+            'width: 80px; '
+            'height: 3px; '
+            'margin: 1rem auto 0; '
+            'background: linear-gradient(90deg, transparent, #6366f1, transparent); '
+            'border-radius: 2px;'
+            '"></div>'
+            '</div>'
+        )
 		with gr.Row(equal_height=True):
 			lang_choice = gr.Radio(
 				choices=["中文", "English"],
@@ -315,7 +336,6 @@ def render_interface() -> gr.Blocks:
 				interactive=True,
 			)
 
-		title_md = gr.Markdown(_i18n("title"))
 		usage_md = gr.Markdown(_usage_md())
 
 		with gr.Row(equal_height=True):
@@ -402,7 +422,6 @@ def render_interface() -> gr.Blocks:
 			inputs=[lang_choice],
 			outputs=[
 				lang_choice,
-				title_md,
 				usage_md,
 				prompt_audio,
 				target_audio,