app.py

# from: https://gradio.app/real_time_speech_recognition/

from transformers import pipeline, Wav2Vec2CTCTokenizer, Wav2Vec2ForCTC, Wav2Vec2ProcessorWithLM
import torch
import pyctcdecode
import kenlm
import gradio as gr
import librosa
import os
import time

#Loading the model and the tokenizer
token_key = os.environ.get("HUGGING_FACE_HUB_TOKEN")
#model_name = "unilux/wav2vec-xls-r-Luxembourgish20-with-LM"
model_name = "unilux/wav2vec-xlsr-300m-Luxembourgish-with-LM"
tokenizer = Wav2Vec2CTCTokenizer.from_pretrained(model_name, use_auth_token=token_key)
model = Wav2Vec2ForCTC.from_pretrained(model_name, use_auth_token=token_key)
processor = Wav2Vec2ProcessorWithLM.from_pretrained(model_name, use_auth_token=token_key)

device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)

p = pipeline("automatic-speech-recognition", model=model, tokenizer=tokenizer, feature_extractor=processor.feature_extractor, decoder=processor.decoder, use_auth_token=token_key)

#p = pipeline("automatic-speech-recognition", model=model_name, use_auth_token = token_key)
#p = pipeline("automatic-speech-recognition", model=model_name, use_auth_token = True)

#tokenizer = Wav2Vec2Tokenizer.from_pretrained(model_name)
#model = Wav2Vec2ForCTC.from_pretrained(model_name)

def load_data(input_file):
  
  """ Function for resampling to ensure that the speech input is sampled at 16KHz.
  """
  sampling_rate = 16_000
  #read the file
  speech, sample_rate = librosa.load(input_file, sr=sampling_rate, mono=True)
  speech = librosa.effects.trim(speech, top_db= 10)
  return speech[0]
    
def asr_pipe(input_file):
  load_data(input_file)
  transcription = p(input_file, chunk_length_s=3, stride_length_s=(1, 1))["text"]
  return transcription
    
gr.Interface(asr_pipe,
             inputs = gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Hei kënnt Dir Är Sprooch iwwert de Mikro ophuelen"),
             outputs = gr.outputs.Textbox(label="Erkannten Text"),
             title="Sproocherkennung fir d'Lëtzebuergescht @uni.lu",
             description = "Dës App convertéiert Är geschwate Sprooch an de (méi oder manner richtegen ;-)) Text!",
             examples = [["ChamberMeisch.wav"], ["Chamber_Fayot_2005.wav"], ["Erlieft-a-Verzielt.wav"], ["Schnessen_Beispill.wav"]], theme="default").launch()