app.py

import streamlit as st
from transformers import (
    AutoModelForSpeechSeq2Seq, 
    AutoProcessor, 
    AutoModelForSeq2SeqLM, 
    AutoTokenizer,
    GenerationConfig
)
import torch
import librosa
import numpy as np
import warnings
import os

warnings.filterwarnings("ignore")

# --- Page Configuration ---
st.set_page_config(page_title="Tunuh ASR and MT System", page_icon="🎙️", layout="wide")

# --- Initialize Session State for Resetting ---
if "reset_key" not in st.session_state:
    st.session_state.reset_key = 0

def clear_callback():
    st.session_state.reset_key += 1

# --- Customized Theme ---
st.markdown("""
    <style>
    .main { background-color: #ffffff; }
    .hero-banner {
        background: linear-gradient(135deg, #800000 0%, #4a0000 100%);
        border-radius: 15px;
        padding: 40px;
        text-align: center;
        margin-bottom: 30px;
        box-shadow: 0 4px 15px rgba(0,0,0,0.3);
    }
    .hero-banner h1 { color: white !important; font-weight: 800; margin-bottom: 10px; }
    .hero-banner p { color: #f5f5f5 !important; font-size: 1.2rem; }
    
    .stAudioInput, .stFileUploader {
        border: 2px solid #800000;
        border-radius: 12px;
        padding: 10px;
        background-color: #fff5f5;
    }
    /* --- ENHANCED FIX FOR BUTTON COLORS --- */
    .process-btn button, .clear-btn button {
        background-color: #800000 !important;
        color: white !important;
        border: 2px solid #800000 !important;
        border-radius: 8px !important;
        height: 55px !important;
        width: 100% !important;
        font-weight: bold !important;
        font-size: 16px !important;
        text-transform: uppercase !important;
    }
    .process-btn button:hover, .clear-btn button:hover {
        background-color: #a52a2a !important;
        border-color: #a52a2a !important;
        color: white !important;
    }
    .process-btn button p, .clear-btn button p {
        color: white !important;
    }
    .process-btn button:active, .clear-btn button:active,
    .process-btn button:focus:not(:active), .clear-btn button:focus:not(:active) {
        background-color: #800000 !important;
        color: white !important;
        border-color: #800000 !important;
        box-shadow: none !important;
    }
    .stTextArea textarea {
        background-color: #f8f9fa;
        border: 1px solid #800000;
    }
    </style>
    """, unsafe_allow_html=True)

# --- Loading Models  ---
@st.cache_resource
def load_asr_model():
    MODEL_PATH = "MaryWambo/Microsoft2-178h"
    device = "cpu"
    
    processor = AutoProcessor.from_pretrained(MODEL_PATH)
    model = AutoModelForSpeechSeq2Seq.from_pretrained(
        MODEL_PATH, 
        torch_dtype=torch.float32, 
        low_cpu_mem_usage=True
    ).to(device)
    return processor, model, device

@st.cache_resource
def load_translation_model():
    MODEL_ID = "facebook/nllb-200-distilled-600M"
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
    model = AutoModelForSeq2SeqLM.from_pretrained(
        MODEL_ID,
        torch_dtype=torch.float32,
        low_cpu_mem_usage=True
    ).to("cpu")
    return model, tokenizer

asr_processor, asr_model, asr_device = load_asr_model()
trans_model, trans_tokenizer = load_translation_model()

# --- Logic Functions ---
def transcribe(audio_file):
    try:
        speech_array, sr = librosa.load(audio_file, sr=16000)
        
        # Extract features using the new processor
        inputs = asr_processor(speech_array, sampling_rate=sr, return_tensors="pt")
        input_features = inputs.input_features.to(asr_device)
        
        # Setup generation structure for the updated model
        gen_config = GenerationConfig.from_model_config(asr_model.config)
        gen_config.update(max_new_tokens=255)
        
        with torch.no_grad():
            predicted_ids = asr_model.generate(input_features, generation_config=gen_config)
            
        return asr_processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
    except Exception as e:
        return f"ASR Error: {str(e)}"

def translate_text(text):
    if not text or "Error" in text: return ""
    trans_tokenizer.src_lang = "kik_Latn"
    inputs = trans_tokenizer(text, return_tensors="pt").to("cpu")
    with torch.no_grad():
        forced_bos_id = trans_tokenizer.convert_tokens_to_ids("eng_Latn")
        outputs = trans_model.generate(**inputs, forced_bos_token_id=forced_bos_id, max_new_tokens=128)
    return trans_tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]

# --- UI Layout ---
st.markdown('<div class="hero-banner"><h1>Tunuh AI System</h1><p>Powering the future of low resource languages through speech technologies</p></div>', unsafe_allow_html=True)

col_input, col_output = st.columns([1, 1], gap="large")

with col_input:
    st.markdown("### 🎙️ Input Speech")
    input_method = st.radio("Select input method", ["Record Voice", "Upload File"], horizontal=True)
    
    audio_data = None
    if input_method == "Record Voice":
        audio_data = st.audio_input("Record", key=f"audio_in_{st.session_state.reset_key}")
    else:
        audio_data = st.file_uploader("Upload audio file", type=["wav", "mp3", "webm", "m4a"], key=f"file_up_{st.session_state.reset_key}")
    
    btn_col1, btn_col2 = st.columns(2)
    with btn_col1:
        st.markdown('<div class="process-btn">', unsafe_allow_html=True)
        run_btn = st.button("PROCESS AUDIO")
        st.markdown('</div>', unsafe_allow_html=True)
    
    with btn_col2:
        st.markdown('<div class="clear-btn">', unsafe_allow_html=True)
        st.button(" CLEAR ", on_click=clear_callback)
        st.markdown('</div>', unsafe_allow_html=True)

with col_output:
    st.markdown("### Output")
    transcript_container = st.empty()
    translation_container = st.empty()

    if run_btn and audio_data:
        with st.spinner("Transcribing..."):
            kikuyu_text = transcribe(audio_data)
        transcript_container.text_area("The transcript", value=kikuyu_text, height=150)

        with st.spinner("Translating..."):
            translated_text = translate_text(kikuyu_text)
        translation_container.text_area("Translated text", value=translated_text, height=150)
    else:
        transcript_container.text_area("The transcript", value="", height=150, placeholder="transcription results")
        translation_container.text_area("Translated text", value="", height=150, placeholder="Translation results")

st.divider()