xvectors

app.py

@@ -10,17 +10,10 @@ processor = SpeechT5Processor.from_pretrained(model_name)
 model = SpeechT5ForTextToSpeech.from_pretrained("emirhanbilgic/speecht5_finetuned_emirhan_tr")
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
 
-# Load the Turkish dataset
-turkish_dataset = load_dataset("erenfazlioglu/turkishvoicedataset", split="train")
+# Load speaker embeddings (using the same as before, as it's not clear if a specific embedding is needed for the Turkish model)
+embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
-# Get an example text and its corresponding audio
-example_item = turkish_dataset[0]
-example_text = example_item['text']
-example_audio = example_item['audio']['array']
-
-# Create speaker embedding from the example audio
-with torch.no_grad():
-    speaker_embeddings = model.get_speaker_embeddings(torch.tensor(example_audio).unsqueeze(0))
 
 def text_to_speech(text):
     inputs = processor(text=text, return_tensors="pt")
@@ -31,7 +24,7 @@ def text_to_speech(text):
 # Create Gradio interface
 iface = gr.Interface(
     fn=text_to_speech,
-    inputs=gr.Textbox(label="Enter Turkish text to convert to speech", value=example_text),
+    inputs=gr.Textbox(label="Enter Turkish text to convert to speech", value="Yapay zekayı seviyorum."),
     outputs=gr.Audio(label="Generated Speech"),
     title="Turkish SpeechT5 Text-to-Speech Demo",
     description="Enter Turkish text and listen to the generated speech using the fine-tuned SpeechT5 model."