app.py

from transformers import pipeline
import gradio as gr
import re
import torch
from pyctcdecode import BeamSearchDecoderCTC
import torch
import librosa
import time


lmID = "aware-ai/german-lowercase-4gram-kenlm"
decoder = BeamSearchDecoderCTC.load_from_hf_hub(lmID)
p = pipeline("automatic-speech-recognition", model="aware-ai/robust-wav2vec2-base-german-lowercase", decoder=decoder)
ttp = pipeline("text2text-generation", model="aware-ai/marian-german-grammar")

def transcribe(audio):
    transcribed = p(audio[1], chunk_length_s=20, stride_length_s=(0, 0))["text"]  
    return transcribed
    
def punctuate(text):
    punctuated = ttp(text, max_length = 512)[0]["generated_text"]
    return punctuated 

def get_asr_interface():
    return gr.Interface(
        fn=transcribe, 
        inputs=[
            gr.inputs.Audio(source="microphone")
        ],
        outputs=[
            "textbox",
        ])
        
def get_punctuate_interface():
    return gr.Interface(
        fn=punctuate, 
        inputs=[
            "textbox"
        ],
        outputs=[
            "textbox",
        ])

interfaces = [
    get_asr_interface(),
    get_punctuate_interface(),
]

gradio.Series(interfaces).launch(server_name = "0.0.0.0")