Update app.py

app.py

@@ -14,25 +14,59 @@ def _grab_best_device(use_gpu=True):
 
 device = _grab_best_device()
 
-HUB_PATH = "ylacombe/vits_vctk_welsh_male"
+default_model_per_language = {
+    "english": "kakao-enterprise/vits-ljs",
+    "spanish": "facebook/mms-tts-spa",
+    "tamil": "facebook/mms-tts-tam"
+}
+
+models_per_language = {
+    "english": [
+        "ylacombe/vits_ljs_irish_male_monospeaker",
+        "ylacombe/vits_ljs_welsh_female_monospeaker",
+        "ylacombe/vits_ljs_welsh_male_monospeaker",
+        "ylacombe/vits_ljs_welsh_female_monospeaker_2",
+        "ylacombe/vits_ljs_welsh_male_monospeaker_2",
+        "ylacombe/vits_ljs_welsh_female_2",
+        "ylacombe/vits_ljs_irish_male_2",
+        "ylacombe/vits_ljs_scottish_female_2",
+    ],
+    "spanish": [
+        "ylacombe/mms-spa-finetuned-monospeaker-all",
+        "ylacombe/mms-spa-finetuned-monospeaker",       
+    ],
+    "tamil": [
+        "ylacombe/mms-tam-finetuned-monospeaker-all",
+        "ylacombe/mms-tam-finetuned-monospeaker",
+    ]
+}
+
+HUB_PATH = "ylacombe/vits_ljs_welsh_female_monospeaker"
 
 
 pipe_dict = {
-    "current_model": "ylacombe/vits_ljs_welsh_male",
+    "current_model": "ylacombe/vits_ljs_welsh_female_monospeaker",
     "pipe":  pipeline("text-to-speech", model=HUB_PATH, device=0),
+    "original_pipe": pipeline("text-to-speech", model=default_model_per_language["english"], device=0),
+    "language": "english",
 }
 
 title = "# 🐶 VITS"
 
+max_speakers = 15
+
 description = """
 
 """
 
-max_speakers = 15
-
 
 # Inference
-def generate_audio(text, model_id):
+def generate_audio(text, model_id, language):
+
+    if pipe_dict["language"] != language:
+        gr.Warning(f"Language has changed - loading new default model: {default_model_per_language[language]}")
+        pipe_dict["language"] == language
+        pipe_dict["original_pipe"] = pipeline("text-to-speech", model=default_model_per_language[language], device=0)
     
     if pipe_dict["current_model"] != model_id:
         gr.Warning("Model has changed - loading new model")
@@ -42,8 +76,15 @@ def generate_audio(text, model_id):
     num_speakers = pipe_dict["pipe"].model.config.num_speakers
 
     out = []
+    # first generate original model result
+    output = pipe_dict["original_pipe"](text)
+    output =  gr.Audio(value = (output["sampling_rate"], output["audio"].squeeze()), type="numpy", autoplay=False, label=f"Non finetuned model prediction {default_model_per_language[language]}", show_label=True,
+                               visible=True)
+    out.append(output)
+    
+    
     if num_speakers>1:
-        for i in range(min(num_speakers, max_speakers)):
+        for i in range(min(num_speakers, max_speakers - 1)):
             forward_params = {"speaker_id": i}
             output = pipe_dict["pipe"](text, forward_params=forward_params)
             
@@ -56,7 +97,7 @@ def generate_audio(text, model_id):
         output =  gr.Audio(value = (output["sampling_rate"], output["audio"].squeeze()), type="numpy", autoplay=False, label="Generated Audio - Mono speaker", show_label=True,
                                visible=True)
         out.append(output)
-        out.extend([gr.Audio(visible=False)]*(max_speakers-1))
+        out.extend([gr.Audio(visible=False)]*(max_speakers-2))
     return out
 
 
@@ -66,25 +107,18 @@ with gr.Blocks() as demo_blocks:
     gr.Markdown(description)
     with gr.Row():
         with gr.Column():
-            inp_text = gr.Textbox(label="Input Text", info="What would you like bark to synthesise?")
+            inp_text = gr.Textbox(label="Input Text", info="What would you like VITS to synthesise?")
             btn = gr.Button("Generate Audio!")
+            language = gr.Dropdown(
+                default_model_per_language.keys(),
+                value = "english",
+                label = "language",
+                info = "Language that you want to test"
+            )
+            
             model_id = gr.Dropdown(
-                    [
-                        "ylacombe/vits_ljs_irish_male_monospeaker",
-                        "ylacombe/vits_ljs_welsh_female_monospeaker",
-                        "ylacombe/vits_ljs_welsh_male_monospeaker",
-                        "ylacombe/vits_ljs_welsh_female_monospeaker_2",
-                        "ylacombe/vits_ljs_welsh_male_monospeaker_2",
-                        "kakao-enterprise/vits-ljs",
-                        "ylacombe/vits_ljs_welsh_female_2",
-                        "ylacombe/vits_ljs_irish_male_2",
-                        "ylacombe/vits_ljs_scottish_female_2",
-                        "ylacombe/mms-tam-finetuned-monospeaker-all",
-                        "ylacombe/mms-spa-finetuned-chilean-monospeaker-all",
-                        "ylacombe/mms-tam-finetuned-monospeaker",
-                        "ylacombe/mms-spa-finetuned-chilean-monospeaker"
-                    ],
-                    value="ylacombe/vits_ljs_welsh_male_2", 
+                    models_per_language["english"],
+                    value="ylacombe/vits_ljs_welsh_female_monospeaker_2", 
                     label="Model", 
                     info="Model you want to test",
                     )
@@ -94,8 +128,18 @@ with gr.Blocks() as demo_blocks:
             for i in range(max_speakers):
                 out_audio = gr.Audio(type="numpy", autoplay=False, label=f"Generated Audio - speaker {i}", show_label=True, visible=False)
                 outputs.append(out_audio)
+
+    language.change(lambda language: gr.Dropdown(
+                    models_per_language[language],
+                    value=models_per_language[language][0], 
+                    label="Model", 
+                    info="Model you want to test",
+                    ),
+                    language,
+                    model_id
+                   )
     
-    btn.click(generate_audio, [inp_text, model_id], outputs)
+    btn.click(generate_audio, [inp_text, model_id, language], outputs)
     
 
 demo_blocks.queue().launch()