app.py

import gradio as gr
import torch
import librosa
import json
from transformers import pipeline
from stitched_model import CombinedModel

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

model = CombinedModel("indonesian-nlp/wav2vec2-luganda", "Sunbird/sunbird-mul-en-mbart-merged", device=device)

def transcribe(audio_file_mic=None, audio_file_upload=None):
    if audio_file_mic:
        audio_file = audio_file_mic
    elif audio_file_upload:
        audio_file = audio_file_upload
    else:
        return "Please upload an audio file or record one"

    # Load the audio file
    speech, sample_rate = librosa.load(audio_file, sr=16000, mono=True)

    # Split the audio into 30-second chunks
    chunk_size = 30 * 16000
    chunks = [speech[i:i + chunk_size] for i in range(0, len(speech), chunk_size)]

    # Process each chunk and concatenate the results
    transcriptions = []
    translations = []
    for chunk in chunks:
        chunk = torch.tensor([chunk])
        with torch.no_grad():
            transcription, translation = model({"audio": chunk})
        transcriptions.append(transcription)
        translations.append(translation[0])

    transcription = "".join(transcriptions)
    translation = "".join(translations)

    return transcription, translation

description = '''Luganda to English Speech Translation'''

iface = gr.Interface(fn=transcribe,
                     inputs=[
                         gr.Audio(source="microphone", type="filepath", label="Record Audio"),
                         gr.Audio(source="upload", type="filepath", label="Upload Audio")],
                     outputs=[gr.Textbox(label="Transcription"),
                              gr.Textbox(label="Translation")
                     ],
                     description=description
                     )
iface.launch()