import gradio as gr
import numpy as np
import torch
from datasets import load_dataset

from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline

speaker_embedding_path = "./speaker_embedding.npy"

replacements = [
    ("&", "og"),
    ("\r", " "),
    ("´", ""),
    ("\\", ""),
    ("¨", " "),
    ("Å", "AA"),
    ("Æ", "AE"),
    ("É", "E"),
    ("Ö", "OE"),
    ("Ø", "OE"),
    ("á", "a"),
    ("ä", "ae"),
    ("å", "aa"),
    ("è", "e"),
    ("î", "i"),
    ("ô", "oe"),
    ("ö", "oe"),
    ("ø", "oe"),
    ("ü", "y"),
]


def replace_danish_letters(text):
    for src, dst in replacements:
        text = text.replace(src, dst)
    return text

device = "cuda:0" if torch.cuda.is_available() else "cpu"

# load speech translation checkpoint
asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)

# load text-to-speech checkpoint and speaker embeddings
processor = SpeechT5Processor.from_pretrained("JackismyShephard/speecht5_tts-finetuned-nst-da")

model = SpeechT5ForTextToSpeech.from_pretrained("JackismyShephard/speecht5_tts-finetuned-nst-da").to(device)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)

speaker_embedding = np.load(speaker_embedding_path)
speaker_embeddings = torch.tensor(speaker_embedding).unsqueeze(0)


def translate(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "da"})
    return outputs["text"]


def synthesise(text):
    inputs = processor(text=text, return_tensors="pt")
    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
    return speech.cpu()


def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    translated_text = replace_danish_letters(translated_text)
    print(translated_text)
    synthesised_speech = synthesise(translated_text)
    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
    return 16000, synthesised_speech


title = "Cascaded STST"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model fine-tuned by [JackismyShephard](https://huggingface.co/JackismyShephard) for Danish for text-to-speech:
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
"""

demo = gr.Blocks()

mic_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(sources=["microphone"], type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    title=title,
    description=description,
)

file_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(sources=["upload"], type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    examples=[["./example.wav"]],
    title=title,
    description=description,
)

with demo:
    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])

demo.launch()