app.py

import streamlit as st
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torchaudio
from io import BytesIO

# Load the model
@st.cache_resource
def load_model():
    processor = WhisperProcessor.from_pretrained("233-Yorozuya/dl_twi_asr")
    model = WhisperForConditionalGeneration.from_pretrained("233-Yorozuya/dl_twi_asr")
    return processor, model

processor, model = load_model()

st.title("ASR with Fine-Tuned Whisper")
st.write("Upload an audio file for transcription:")

# File upload
audio_file = st.file_uploader("Choose an audio file", type=["wav", "mp3", "ogg"])

if audio_file:
    try:
        # Convert uploaded file to bytes
        audio_bytes = BytesIO(audio_file.read())
        audio, rate = torchaudio.load(audio_bytes)
        audio = torchaudio.transforms.Resample(orig_freq=rate, new_freq=16000)(audio)

        # Preprocess the audio
        inputs = processor(audio[0].numpy(), sampling_rate=16000, return_tensors="pt")

        # Specify the language (Asanti Twi)
        model.config.forced_decoder_ids = None  # Disable forced language


        # Perform inference
        with st.spinner("Transcribing..."):
            predicted_ids = model.generate(inputs.input_features)
            transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

        # Display result
        st.subheader("Transcription")
        st.write(transcription)
    except Exception as e:
        st.error(f"An error occurred: {e}")