streamlit_app.py

import streamlit as st
import whisper
import tempfile
import os
import re
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# =========================
# PAGE CONFIG
# =========================
st.set_page_config(page_title="Speech to Text Translator", page_icon="🎙️", layout="centered")

# =========================
# UI STYLE
# =========================
st.markdown("""
<style>
html, body, [data-testid="stAppViewContainer"] {
    background: radial-gradient(circle at 20% 20%, #1e293b, #020617 70%);
    color: white;
    font-family: 'Inter', sans-serif;
}
.title {
    text-align: center;
    font-size: 40px;
    font-weight: 700;
}
.subtitle {
    text-align: center;
    color: #cbd5e1;
    margin-bottom: 22px;
}
[data-testid="stFileUploader"] {
    border-radius: 16px;
    border: 1px dashed rgba(255,255,255,0.25);
}
.result-box {
    background: rgba(16,185,129,0.12);
    border-radius: 16px;
    padding: 16px;
    border: 1px solid rgba(16,185,129,0.35);
}
</style>
""", unsafe_allow_html=True)

# =========================
# HEADER
# =========================
st.markdown('<div class="title"> Speech to Text Translator</div>', unsafe_allow_html=True)
st.markdown('<div class="subtitle">Transcribe speech or translate into any language</div>', unsafe_allow_html=True)

# =========================
# TEXT UTILS
# =========================
def split_text(text, max_len=200):
    sentences = re.split(r'(?<=[.!?。！？])', text)
    chunks = []
    cur = ""
    for s in sentences:
        if len(cur) + len(s) < max_len:
            cur += " " + s
        else:
            chunks.append(cur.strip())
            cur = s
    if cur:
        chunks.append(cur.strip())
    return chunks

# =========================
# MODEL CACHE
# =========================
@st.cache_resource
def load_models():
    whisper_model = whisper.load_model("base")
    model_name = "facebook/nllb-200-distilled-600M"
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    nllb_model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
    return whisper_model, tokenizer, nllb_model

whisper_model, tokenizer, nllb_model = load_models()

# =========================
# LANGUAGE MAP
# =========================
LANG_CODE = {
    "Arabic":"arb_Arab","Assamese":"asm_Beng","Awadhi":"awa_Deva",
    "Bengali":"ben_Beng","Bhojpuri":"bho_Deva","Chinese":"zho_Hans",
    "English":"eng_Latn","French":"fra_Latn","German":"deu_Latn",
    "Hindi":"hin_Deva","Japanese":"jpn_Jpan","Korean":"kor_Hang",
    "Maithili":"mai_Deva","Marathi":"mar_Deva","Persian":"pes_Arab",
    "Punjabi":"pan_Guru","Russian":"rus_Cyrl","Sanskrit":"san_Deva",
    "Spanish":"spa_Latn","Tamil":"tam_Taml","Telugu":"tel_Telu",
    "Urdu":"urd_Arab","Vietnamese":"vie_Latn"
}

# =========================
# OPTIONS
# =========================
col1, col2 = st.columns(2)
with col1:
    transcribe = st.checkbox("📄 Transcribe")
with col2:
    translate = st.checkbox("🌍 Translate", value=True)

target_lang = None
if translate:
    target_lang = st.selectbox("Translate into", sorted(LANG_CODE.keys()))

# =========================
# UPLOAD
# =========================
audio_file = st.file_uploader(
    "Upload audio (MP3, WAV, M4A, MP4)",
    type=["mp3","wav","m4a","mp4"]
)

if audio_file:
    st.audio(audio_file)

# =========================
# PROCESS
# =========================
if st.button(" Process Audio") and audio_file:

    with st.spinner("Processing audio..."):

        # Save uploaded audio
        with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp:
            tmp.write(audio_file.read())
            tmp_path = tmp.name

        # ---------- TRANSCRIBE ----------
        result = whisper_model.transcribe(tmp_path, fp16=False)
        detected_lang = result["language"]
        text = result["text"]

        st.success(f"Detected language: {detected_lang}")

        output_text = text

        # ---------- TRANSLATE ----------
        if translate and target_lang:
            tgt_code = LANG_CODE[target_lang]
            chunks = split_text(text)

            translated_parts = []

            for chunk in chunks:
                inputs = tokenizer(chunk, return_tensors="pt")

                tokens = nllb_model.generate(
                    **inputs,
                    forced_bos_token_id=tokenizer.convert_tokens_to_ids(tgt_code),
                    max_length=256,
                    num_beams=4,
                    no_repeat_ngram_size=3,
                    repetition_penalty=1.2,
                    early_stopping=True
                )

                translated = tokenizer.batch_decode(tokens, skip_special_tokens=True)[0]
                translated_parts.append(translated)

            output_text = " ".join(translated_parts)

        # ================= OUTPUT =================
        st.markdown("### 📝 Output Text")
        st.markdown(f'<div class="result-box">{output_text}</div>', unsafe_allow_html=True)

        st.download_button(
            "⬇ Download Text",
            output_text,
            file_name="translated_text.txt"
        )

        os.remove(tmp_path)