app.py

import gradio as gr
import time
from transformers import pipeline
import torch
import ffmpeg  # Make sure it's ffmpeg-python

# Check if GPU is available
use_gpu = torch.cuda.is_available()


# Configure the pipeline to use the GPU if available
if use_gpu:
    p = pipeline("automatic-speech-recognition", 
             model="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-faroese-100h", device=0)
else:
    p = pipeline("automatic-speech-recognition", 
             model="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-faroese-100h")


def extract_audio_from_m3u8(url):
    try:
        output_file = "output_audio.aac"
        ffmpeg.input(url).output(output_file).run(overwrite_output=True)
        return output_file
    except Exception as e:
        return f"An error occurred: {e}"


def transcribe(audio, state="", uploaded_audio=None, m3u8_url=""):
    if m3u8_url:
        audio = extract_audio_from_m3u8(m3u8_url)

    if uploaded_audio is not None:
        audio = uploaded_audio

    if not audio:
        return state, state  # Return a meaningful message

    try:
        time.sleep(3)
        text = p(audio, chunk_length_s= 50)["text"]
        state += text + "\n"
        return state, state
    except Exception as e:
        return "An error occurred during transcription.", state  # Handle other exceptions


def reset(state):
    state = ''
    return state


demo = gr.Interface(
    fn=transcribe, 
    inputs=[
        gr.components.Audio(source="microphone", type="filepath"),
        'state',
        gr.components.Audio(label="Upload Audio File", type="filepath", source="upload"),
        gr.components.Textbox(label="m3u8 URL | E.g.: from kvf.fo or logting.fo")
    ],
    outputs=[
        gr.components.Textbox(type="text"),
        "state"
    ],

    live=True)


demo.launch()