app.py

import streamlit as st
import librosa
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
import tempfile
from pydub import AudioSegment

# Define available models
available_models = ["Yehor/whisper-small-ukrainian", "arampacha/whisper-large-uk-2"]
# , "Yehor/wav2vec2-xls-r-300m-uk-with-3gram-news-lm", "Yehor/wav2vec2-xls-r-300m-uk-with-wiki-lm"
# available_models = ["theodotus/stt_uk_squeezeformer_ctc_sm", "arampacha/whisper-large-uk-2"]


st.title("Voice Recognition App")

# Model selection dropdown
model_choice = st.selectbox("Choose a model", available_models)

processor = AutoProcessor.from_pretrained(model_choice)

model = AutoModelForSpeechSeq2Seq.from_pretrained(model_choice)

uploaded_file = st.file_uploader("Choose file", type=["wav", "mp3"])

def map_to_pred(file_path):
    # load audio file
    audio, _ = librosa.load(file_path)

    # preprocess audio and generate standard
    input_features = processor([audio], return_tensors="pt", sampling_rate=16000).input_features
    generated_ids = model.generate(inputs=input_features)
    transcription = processor.batch_decode(generated_ids, normalize=True, skip_special_tokens=True)
    text = processor.tokenizer._normalize(transcription[0])

    return text
if uploaded_file is not None:
    # convert file object to file path
    file_path = './temp.wav'
    with open(file_path, 'wb') as f:
        f.write(uploaded_file.getbuffer())

    # Save the uploaded file temporarily
    with tempfile.NamedTemporaryFile(delete=False) as temp_file:
        temp_file.write(uploaded_file.read())
        temp_file_path = temp_file.name

    # Convert audio file to a format supported by Whisper (if necessary)
    audio = AudioSegment.from_file(temp_file_path)
    temp_wav_path = tempfile.mktemp(suffix=".wav")
    audio.export(temp_wav_path, format="wav")

    st.audio(uploaded_file, format="audio/wav")

    text = map_to_pred(file_path)

    # display results
    st.write('Input audio:', uploaded_file.name)
    st.write('Predicted standard:', text)