import gradio as gr
import numpy as np
import torch
from transformers import BarkModel
from transformers import AutoProcessor
from transformers import pipeline
import librosa

processor = AutoProcessor.from_pretrained("suno/bark-small")
model = BarkModel.from_pretrained("suno/bark-small")


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


# https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c
language_presets = {"es":"v2/es_speaker_",
                    "en":"v2/en_speaker_"}
def tts(text, language="es", style:int = 0):
    voice_preset = language_presets[language] + str(style)
    # prepare the inputs
    inputs = processor(text, voice_preset = voice_preset)
    # generate speech
    speech_output = model.generate(**inputs.to(device))
    sampling_rate = model.generation_config.sample_rate
    return speech_output[0].cpu().numpy(), sampling_rate


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


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

    return text


def synthesise(text, language="es",style=0):
    speech, sr = tts(text, language=language, style=style)
    target_sr = 16_000
    speech = librosa.resample(speech, orig_sr = sr, target_sr = target_sr)
    return speech, target_sr


def speech_to_speech_translation(audio, debug = True):
    translated_text = translate(audio)
    if debug:
        print(f"{translated_text=}")
    synthesised_speech, sampling_rate = synthesise(translated_text)
    # tranform to int for Gradio
    synthesised_speech = (np.array(synthesised_speech) * 32767).astype(np.int16)
    return sampling_rate, 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 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(source="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(source="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()