app.py

import gradio as gr
import whisper
from transformers import pipeline

# Load Whisper model
whisper_model = whisper.load_model("small", device="cpu")

# Load the text correction model
correction_pipeline = pipeline("text2text-generation", model="tiiuae/falcon3-1b-instruct", device=-1)

# Function to preprocess audio and transcribe it using Whisper
def transcribe_audio(audio_file):
    transcription = whisper_model.transcribe(audio_file)
    return transcription["text"]

# Function to correct grammar in text
def chunk_text(text, max_tokens=2000):
    """
    Splits the text into smaller chunks to ensure it doesn't exceed the token limit.
    """
    words = text.split()
    chunks = []
    chunk = []
    current_tokens = 0

    for word in words:
        word_tokens = len(word.split())  # Approximate token count
        if current_tokens + word_tokens > max_tokens:
            chunks.append(" ".join(chunk))
            chunk = [word]
            current_tokens = word_tokens
        else:
            chunk.append(word)
            current_tokens += word_tokens

    if chunk:
        chunks.append(" ".join(chunk))

    return chunks


# Function to process the pipeline
def process_pipeline(audio_file):
    raw_transcription = transcribe_audio(audio_file)
    corrected_transcription = correct_text(raw_transcription)
    return raw_transcription, corrected_transcription

# Gradio Interface
interface = gr.Interface(
    fn=process_pipeline,
    inputs=gr.Audio(type="filepath", label="Upload Audio"),
    outputs=[
        gr.Textbox(label="Raw Transcription"),
        gr.Textbox(label="Corrected Transcription"),
    ],
    title="Speech Correction Demo",
    description="Upload an audio file to see raw transcription and grammar-corrected output.",
)

# Launch the app
interface.launch(share=True)