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("ak3ra/wav2vec2-sunbird-speech-lug", "Sunbird/sunbird-mul-en-mbart-merged", device="cpu")


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"

    # Make sure audio is 16kHz
    speech, sample_rate = librosa.load(audio_file)
    if sample_rate != 16000:
        speech = librosa.resample(speech, orig_sr=sample_rate, target_sr=16000)
    speech = torch.tensor([speech])

    with torch.no_grad():
        transcription, translation = model({"audio":speech})
    
    return transcription, translation[0]

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()