import torch

from transformers import pipeline

import numpy as np
import gradio as gr

def _grab_best_device(use_gpu=True):
    if torch.cuda.device_count() > 0 and use_gpu:
        device = "cuda"
    else:
        device = "cpu"
    return device

device = _grab_best_device()

default_model_per_language = {
    "english": "kakao-enterprise/vits-ljs",
    "spanish": "facebook/mms-tts-spa",
}

models_per_language = {
    "english": [
        ("Irish Male Speaker", "ylacombe/vits_ljs_irish_male_monospeaker_2"),
        ("Welsh Female Speaker", "ylacombe/vits_ljs_welsh_female_monospeaker_2"),
        ("Welsh Male Speaker", "ylacombe/vits_ljs_welsh_male_monospeaker_2"),
        ("Scottish Female Speaker", "ylacombe/vits_ljs_scottish_female_monospeaker"),
        ("Northern Female Speaker", "ylacombe/vits_ljs_northern_female_monospeaker"),
        ("Midlands Male Speaker", "ylacombe/vits_ljs_midlands_male_monospeaker"),
        ("Southern Male Speaker", "ylacombe/vits_ljs_southern_male_monospeaker"),
    ],
    "spanish": [
        ("Male Chilean Speaker", "ylacombe/mms-spa-finetuned-chilean-monospeaker"),       
        ("Female Argentinian Speaker", "ylacombe/mms-spa-finetuned-argentinian-monospeaker"),       
        ("Male Colombian Speaker", "ylacombe/mms-spa-finetuned-colombian-monospeaker"),       
    ],
}

pipe_dict = {
    "pipe":  [pipeline("text-to-speech", model=l[1], device=0) for l in models_per_language["english"]],
    "original_pipe": pipeline("text-to-speech", model=default_model_per_language["english"], device=0),
    "language": "english",
}

title = "# VITS"

description = """
TODO
"""

max_speakers = 15

# Inference
def generate_audio(text, language):

    if pipe_dict["language"] != language:
        gr.Warning(f"Language has changed - loading corresponding models: {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)
    
        pipe_dict["pipe"] = [pipeline("text-to-speech", model=l[1], device=0) for l in models_per_language[language]]


    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"Prediction from the original checkpoint {default_model_per_language[language]}", show_label=True,
                               visible=True)
    out.append(output)
    
    
    for i in range(min(len(pipe_dict["pipe"]), max_speakers - 1)):
        
        output = pipe_dict["pipe"][i](text)
        
        output =  gr.Audio(value = (output["sampling_rate"], output["audio"].squeeze()), type="numpy", autoplay=False, label=f"Finetuned {models_per_language[language][i][0]}", show_label=True,
                           visible=True)
        out.append(output)
    
    out.extend([gr.Audio(visible=False)]*(max_speakers-(len(out))))
    return out


# Gradio blocks demo    
with gr.Blocks() as demo_blocks:
    gr.Markdown(title)
    gr.Markdown(description)
    with gr.Row():
        with gr.Column():
            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"
            )
            
        with gr.Column():
            outputs = []
            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)
    
    btn.click(generate_audio, [inp_text, language], outputs)
    

demo_blocks.queue().launch()