app.py

import gradio as gr
import librosa
import soundfile as sf
import torch
import warnings
import os 
from transformers import Wav2Vec2ProcessorWithLM, Wav2Vec2CTCTokenizer

warnings.filterwarnings("ignore")

#load wav2vec2 tokenizer and model

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

from fastapi import FastAPI, HTTPException, File

from transformers import pipeline


pipe = pipeline(model="Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2",chunk_length_s=20, stride_length_s=(4, 2))

token = os.getenv()


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_checkpoint = 'Finnish-NLP/t5x-small-nl24-finnish'    
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, use_auth_token=os.environ.get('hf_token'))
model = AutoModelForSeq2SeqLM.from_pretrained('Finnish-NLP/case_correction_model', from_flax=False, torch_dtype=torch.float32, use_auth_token=os.environ.get('hf_token')).to(device)


# define speech-to-text function
def asr_transcript(audio):
    
    text = ""

    if audio:
        text = pipe(audio.name)
        
        input_ids = tokenizer(text['text'], return_tensors="pt").input_ids.to(device)
        outputs = model.generate(input_ids, max_length=128)
        case_corrected_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
        return {"text_asr": text['text'], "text_case_corrected": case_corrected_text}
    else:
        return  "File not valid"
    
gradio_ui = gr.Interface(
    fn=asr_transcript,
    title="Speech-to-Text with HuggingFace+Wav2Vec2",
    description="Upload an audio clip, and let AI do the hard work of transcribing",
    inputs=gr.inputs.Audio(label="Upload Audio File", type="file"),
    outputs=gr.outputs.Textbox(label="Auto-Transcript"),
)

gradio_ui.launch()