import torch
import os
from transformers import pipeline, VitsModel, VitsTokenizer, SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor
import numpy as np
os.system("pip install git+https://github.com/openai/whisper.git")
import gradio as gr
import whisper

model = whisper.load_model("small")
device = "cuda:0" if torch.cuda.is_available() else "cpu"

def inference(audio):
    audio = whisper.load_audio(audio)
    print("loading finished")
    audio = whisper.pad_or_trim(audio)
    print("audio trimed")
    mel = whisper.log_mel_spectrogram(audio).to(model.device)
    print("spectro finished")
    _, probs = model.detect_language(mel)
    print("lang detected")
    options = whisper.DecodingOptions(fp16 = False)
    print("options decoded")
    result = whisper.decode(model, mel, options)
    
    print(result.text)
    return result.text

    
# Load Whisper-small
pipe = pipeline("automatic-speech-recognition",
                model="openai/whisper-small",
                device=device
)

# Load the model checkpoint and tokenizer
#model = VitsModel.from_pretrained("Matthijs/mms-tts-eng")
#tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-eng")
model2 = VitsModel.from_pretrained("facebook/mms-tts-eng")
tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")


# Define a function to translate an audio, in english here
def translate(audio):
    # return inference(audio)
    outputs = pipe(audio, max_new_tokens=256,
                   generate_kwargs={"task": "translate"})
    return outputs["text"]


# Define function to generate the waveform output
def synthesise(text):
    inputs = tokenizer(text, return_tensors="pt")
    input_ids = inputs["input_ids"]

    with torch.no_grad(): 
      outputs = model2(input_ids)
    
    return outputs.audio[0]


# Define the pipeline
def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    synthesised_speech = synthesise(translated_text)
    synthesised_speech = (
        synthesised_speech.numpy() * 32767).astype(np.int16)
    return [translated_text, (16000, synthesised_speech)]

def predict(transType, language, audio, audio_mic = None):
        print("debug1:", audio,"debug2", audio_mic)
        if not audio and audio_mic:
            audio = audio_mic
        
        if transType == "Text":
            return translate(audio), None
        if transType == "Audio":
            return speech_to_speech_translation(audio)
            
# Define the title etc
title = "Swedish STSOT (Speech To Speech Or Text)"
description="Use Whisper pretrained model to convert swedish audio to english (text or audio)"


supportLangs = ["Swedish", "French (in training)"]
transTypes = ["Text", "Audio"]

#examples = [
#    ["Text", "Swedish", "./ex1.wav", None],
#    ["Audio", "Swedish", "./ex2.wav", None]
#]

examples =[]
demo = gr.Interface(
    fn=predict,
    inputs=[
        gr.Radio(label="Choose your output format", choices=transTypes),
        gr.Radio(label="Choose a source language", choices=supportLangs, value="Swedish"),
        gr.Audio(label="Import an audio", sources="upload", type="filepath"),
        #gr.Audio(label="Import an audio", sources="upload", type="numpy"),
        gr.Audio(label="Record an audio", sources="microphone", type="filepath"),
    ],
    outputs=[
        gr.Text(label="Text translation"),gr.Audio(label="Audio translation",type = "numpy")
    ],
    title=title,
    description=description,
    article="",
    examples=examples,
)


demo.launch()