import gradio as gr
from transformers import pipeline, VitsModel, AutoTokenizer
import torch
import numpy as np

# Load the pipeline for speech recognition and translation
pipe = pipeline(
    "automatic-speech-recognition",
    model="Baghdad99/saad-speech-recognition-hausa-audio-to-text",
    tokenizer="Baghdad99/saad-speech-recognition-hausa-audio-to-text"
)
translator = pipeline("text2text-generation", model="Baghdad99/saad-hausa-text-to-english-text")

# Load the VITS model for text-to-speech synthesis
tts_model = VitsModel.from_pretrained("Baghdad99/english_voice_tts")
tts_tokenizer = AutoTokenizer.from_pretrained("Baghdad99/english_voice_tts")

# Define the function to translate speech
def translate_speech(audio):
    # Separate the sample rate and the audio data
    sample_rate, audio_data = audio

    # Use the speech recognition pipeline to transcribe the audio
    output = pipe(audio_data)
    print(f"Output: {output}")  # Print the output to see what it contains

    # Check if the output contains 'text'
    if 'text' in output:
        transcription = output["text"]
    else:
        print("The output does not contain 'text'")
        return

    # Use the translation pipeline to translate the transcription
    translated_text = translator(transcription, return_tensors="pt")
    print(f"Translated text: {translated_text}")  # Print the translated text to see what it contains

    # Check if the translated text contains 'generated_token_ids'
    if 'generated_token_ids' in translated_text[0]:
        # Decode the tokens into text
        translated_text_str = translator.tokenizer.decode(translated_text[0]['generated_token_ids'])
    else:
        print("The translated text does not contain 'generated_token_ids'")
        return

    # Use the VITS model to synthesize the translated text
    tts_inputs = tts_tokenizer(translated_text_str, return_tensors="pt")
    with torch.no_grad():
        synthesised_speech = tts_model(**tts_inputs).waveform
    print(f"Synthesised speech: {synthesised_speech}")  # Print the synthesised speech to see what it contains

    # Define the max_range variable
    max_range = 1.0  # You can adjust this value based on your requirements
    synthesised_speech = (synthesised_speech.numpy() * max_range).astype(np.float32)

    return 16000, synthesised_speech

# Define the Gradio interface
iface = gr.Interface(
    fn=translate_speech, 
    inputs=gr.inputs.Audio(source="microphone", type="numpy"), 
    outputs=gr.outputs.Audio(type="numpy"),
    title="Hausa to English Translation",
    description="Realtime demo for Hausa to English translation using speech recognition and text-to-speech synthesis."
)

iface.launch()