app.py

import torch, gradio as gr, soundfile as sf, tempfile
from transformers import VitsModel, AutoProcessor

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

MODELS = {
    "Male":   "vlntlbr/vits-greek-male",
    "Female": "vlntlbr/vits-greek-female",
}
tts  = {k: VitsModel.from_pretrained(r).to(DEVICE) for k, r in MODELS.items()}
proc = {k: AutoProcessor.from_pretrained(r)        for k, r in MODELS.items()}

SNIPPETS = {
    "Male":   "audio/male_ref.mp3",
    "Female": "audio/female_ref.mp3",
}


def synth(text, speaker):
    inputs = proc[speaker](text, return_tensors="pt").to(DEVICE)
    with torch.no_grad():
        wav = tts[speaker](**inputs).waveform.squeeze().cpu()
    tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".mp3")
    sf.write(tmp.name, wav.numpy(), 16000)
    return tmp.name

# Add examples (Greek phrases with their respective speakers)
examples = [
    ["Πες μου πώς σε λένε.", "Male"],  
    ["Τι μας προτείνεις;", "Female"],
    ["Είναι τόσο όμορφα έξω!", "Male"],  
    ["Η οικογένεια είναι μαζεμένη στην τραπεζαρία", "Female"] 
]

custom_css = """

h1, .gr-markdown h1 {
    text-align: center;
}
.gr-markdown p {
    text-align: center;
}
"""

with gr.Blocks(title="Greek TTS (male / female)", css=custom_css) as demo:
    gr.Markdown(
        "# Greek TTS Demo\n"
        "Choose a speaker, listen to a reference clip, then enter Greek text."
    )

    with gr.Row():
        speaker = gr.Radio(["Male", "Female"], value="Male", label="Speaker")
        ref_aud = gr.Audio(SNIPPETS["Male"], interactive=False, label="Reference")
        speaker.change(lambda s: gr.update(value=SNIPPETS[s]), speaker, ref_aud)

    text_in = gr.Textbox(label="Greek text", placeholder="Γράψε κάτι…")
    
    gr.Examples(examples=examples, inputs=[text_in, speaker])

    out_aud = gr.Audio(label="Synthesised speech")
    gr.Button("Generate!").click(synth, [text_in, speaker], out_aud)

demo.launch()