app.py

import torch  # Add this line
import gradio as gr
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor, pipeline, AutoTokenizer
import numpy as np
import librosa

# Load the models and processors
asr_model = Wav2Vec2ForCTC.from_pretrained("Akashpb13/Hausa_xlsr")
asr_processor = Wav2Vec2Processor.from_pretrained("Akashpb13/Hausa_xlsr")
translator = pipeline("text2text-generation", model="Baghdad99/saad-hausa-text-to-english-text")
tts = pipeline("text-to-speech", model="Baghdad99/english_voice_tts")

def translate_speech(audio_input):
    # Load the audio file as a floating point time series
    audio_data, sample_rate = librosa.load(audio_input, sr=None)

    # Prepare the input dictionary
    input_dict = asr_processor(audio_data, sampling_rate=sample_rate, return_tensors="pt", padding=True)

    # Use the ASR model to get the logits
    logits = asr_model(input_dict.input_values.to("cpu")).logits

    # Get the predicted IDs
    pred_ids = torch.argmax(logits, dim=-1)[0]

    # Decode the predicted IDs to get the transcription
    transcription = asr_processor.decode(pred_ids)
    print(f"Transcription: {transcription}")  # Print the transcription

    # 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

    # 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'])
        print(f"Translated text string: {translated_text_str}")  # Print the translated text string
    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)

    # 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(type="filepath"),  # 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()