app.py

import gradio as gr
import librosa
import torch
from transformers import Wav2Vec2CTCTokenizer, Wav2Vec2ForCTC, Wav2Vec2Processor

# Load tokenizer, processor, and model
tokenizer = Wav2Vec2CTCTokenizer("./vocab.json", unk_token="[UNK]", pad_token="[PAD]", word_delimiter_token="|")
processor = Wav2Vec2Processor.from_pretrained('boumehdi/wav2vec2-large-xlsr-moroccan-darija', tokenizer=tokenizer)
model = Wav2Vec2ForCTC.from_pretrained('boumehdi/wav2vec2-large-xlsr-moroccan-darija')

# Define the function for transcribing audio
def transcribe(audio):
    # Load the audio data from the Gradio input (audio is in the format of a NumPy array)
    input_audio = audio
    sr = 16000  # Ensure the sample rate is 16000 Hz, which is expected by the model

    # Tokenize the audio
    input_values = processor(input_audio, return_tensors="pt", padding=True).input_values

    # Get the model's logits
    logits = model(input_values).logits

    # Find the predicted tokens
    tokens = torch.argmax(logits, axis=-1)

    # Decode the tokens to text
    transcription = tokenizer.batch_decode(tokens)

    return transcription[0]

# Create the Gradio interface
interface = gr.Interface(
    fn=transcribe,  # Function to be called when an audio file is uploaded or recorded
    inputs=[
        gr.Audio(source="upload", type="numpy"),  # Allow user to upload an audio file
        gr.Audio(source="microphone", type="numpy")  # Allow user to record audio from the browser
    ],
    outputs="text",  # Output will be a transcription text
    title="Moroccan Darija Speech-to-Text",  # Interface title
    description="Upload an audio file or record audio directly from your microphone to transcribe it into Moroccan Darija."
)

# Launch the interface
interface.launch()