app.py

import gradio as gr
from pathlib import Path
import nemo.collections.asr as nemo_asr
import librosa
import soundfile as sf

base_path = str(Path(__file__).parent)

# Convert audio to 16kHz WAV
def convert_wav_to_16k(input_wav, output_file_path, sr=16000):
    if input_wav is None:
        raise ValueError("No audio file provided")

    if isinstance(input_wav, str):  # filepath
        y, s = librosa.load(input_wav, sr=sr)
    elif isinstance(input_wav, tuple):  # numpy array + sample rate
        y, orig_sr = input_wav
        y = librosa.resample(y, orig_sr=orig_sr, target_sr=sr)
        s = sr
    else:
        raise ValueError(f"Unsupported audio input type: {type(input_wav)}")

    sf.write(output_file_path, y, s)
    print(f'"{output_file_path}" has been converted to {s}Hz')
    return output_file_path

# Load NeMo model and run transcription
def loading_nemo_and_prediction(processed_wav):
    arabic_asr = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from(
        restore_path="FastConformer-Custom-Tokenizer.nemo"
    )
    prediction = arabic_asr.transcribe(paths2audio_files=[processed_wav])
    return prediction

# Prediction wrapper
def predict(uploaded_wav):
    if uploaded_wav is None:
        return "No audio file uploaded."

    out_path = base_path + "/converted.wav"
    audio_conversion = convert_wav_to_16k(uploaded_wav, out_path)
    prediction_text = loading_nemo_and_prediction(audio_conversion)
    return prediction_text[0]

# Gradio interface
demo = gr.Interface(
    fn=predict,
    inputs=gr.Audio(label="Upload or record audio", interactive=True, type="filepath"),
    outputs=gr.Textbox(label="Transcription")
)

demo.launch(debug=True, share=True)