import gradio as gr
from transformers import pipeline, AutoTokenizer
from huggingsound import SpeechRecognitionModel
import numpy as np

# Load the model for speech recognition
model = SpeechRecognitionModel("jonatasgrosman/wav2vec2-large-xlsr-53-english")

translator = pipeline("text2text-generation", model="Baghdad99/saad-english-text-to-hausa-text")
tts = pipeline("text-to-speech", model="Baghdad99/hausa_voice_tts")

# Define the function to translate speech
def translate_speech(audio_data_tuple):
    print(f"Type of audio: {type(audio_data_tuple)}, Value of audio: {audio_data_tuple}")  # Debug line

    # Extract the audio data from the tuple
    sample_rate, audio_data = audio_data_tuple

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

    # Use the translation pipeline to translate the transcription
    translated_text = translator(output, 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 text-to-speech pipeline to synthesize the translated text
    synthesised_speech = tts(translated_text_str)
    print(f"Synthesised speech: {synthesised_speech}")  # Print the synthesised speech to see what it contains

    # Check if the synthesised speech contains 'audio'
    if 'audio' in synthesised_speech:
        synthesised_speech_data = synthesised_speech['audio']
    else:
        print("The synthesised speech does not contain 'audio'")
        return

    # Flatten the audio data
    synthesised_speech_data = synthesised_speech_data.flatten()

    # Scale the audio data to the range of int16 format
    synthesised_speech = (synthesised_speech_data * 32767).astype(np.int16)

    return 16000, synthesised_speech

# Define the Gradio interface
iface = gr.Interface(
    fn=translate_speech, 
    inputs=gr.inputs.Audio(source="microphone"),  # Change this line
    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()