app.py

import gradio as gr
import re
from transformers import MarianMTModel, MarianTokenizer

# -------------------------
# Fast Model
# -------------------------
MODEL_NAME = "Helsinki-NLP/opus-mt-en-hi"

tokenizer = MarianTokenizer.from_pretrained(MODEL_NAME)

model = MarianMTModel.from_pretrained(MODEL_NAME)

# -------------------------
# Parse SRT
# -------------------------
def parse_srt(srt_text):

    pattern = re.compile(
        r"(\d+)\s*\n"
        r"(\d{2}:\d{2}:\d{2},\d{3}\s-->\s\d{2}:\d{2}:\d{2},\d{3})\s*\n"
        r"(.*?)(?=\n\d+\n|\Z)",
        re.DOTALL
    )

    matches = pattern.findall(srt_text.strip())

    subtitles = []

    for idx, timestamp, text in matches:

        subtitles.append({
            "index": idx,
            "timestamp": timestamp,
            "text": text.replace("\n", " ").strip()
        })

    return subtitles


# -------------------------
# Shorten Hindi
# -------------------------
def shorten_text(text, max_len):

    if len(text) <= max_len:
        return text

    words = text.split()

    while len(text) > max_len and len(words) > 1:
        words.pop()
        text = " ".join(words)

    return text


# -------------------------
# Batch Translate
# -------------------------
def batch_translate(texts):

    inputs = tokenizer(
        texts,
        return_tensors="pt",
        padding=True,
        truncation=True
    )

    translated = model.generate(
        **inputs,
        max_new_tokens=64
    )

    outputs = tokenizer.batch_decode(
        translated,
        skip_special_tokens=True
    )

    return outputs


# -------------------------
# Main Function
# -------------------------
def translate_srt(srt_text):

    subtitles = parse_srt(srt_text)

    english_texts = [
        sub["text"] for sub in subtitles
    ]

    # FAST batch translation
    hindi_texts = batch_translate(
        english_texts
    )

    output = []

    for sub, hindi in zip(subtitles, hindi_texts):

        english_len = len(sub["text"])

        # 130% rule
        max_hindi_len = int(
            english_len * 1.3
        )

        hindi = shorten_text(
            hindi.strip(),
            max_hindi_len
        )

        block = (
            f'{sub["index"]}\n'
            f'{sub["timestamp"]}\n'
            f'{hindi}\n'
        )

        output.append(block)

    return "\n".join(output)


# -------------------------
# UI
# -------------------------
demo = gr.Interface(
    fn=translate_srt,
    inputs=gr.Textbox(
        lines=20,
        label="English SRT"
    ),
    outputs=gr.Textbox(
        lines=20,
        label="Hindi SRT"
    ),
    title="Fast English → Hindi SRT Translator",
    description="Ultra-fast subtitle translation with timestamp preservation and subtitle length control."
)

demo.launch()