app.py

import streamlit as st
import os
import librosa
import torch
from pydub import AudioSegment
from pydub.silence import split_on_silence
from dotenv import load_dotenv
from tempfile import NamedTemporaryFile
import math
from docx import Document
from transformers import WhisperProcessor, WhisperForConditionalGeneration

# Load environment variables from .env file (if needed)
load_dotenv()

@st.cache_resource
def load_whisper_model():
    """
    Load the Whisper model and processor from Hugging Face.
    You can change the model variant ("openai/whisper-base" is used here).
    """
    model_name = "openai/whisper-base"  # Options: "tiny", "base", "small", "medium", "large"
    processor = WhisperProcessor.from_pretrained(model_name)
    model = WhisperForConditionalGeneration.from_pretrained(model_name)
    return processor, model

processor, model = load_whisper_model()

def split_audio_on_silence(audio_file_path, min_silence_len=500, silence_thresh=-40, keep_silence=250):
    """
    Split an audio file into chunks using silence detection.
    
    Args:
        audio_file_path (str): Path to the audio file.
        min_silence_len (int): Minimum length of silence (in ms) required for a split.
        silence_thresh (int): The volume (in dBFS) below which is considered silence.
        keep_silence (int): Amount of silence (in ms) to retain at the beginning and end of each chunk.
    
    Returns:
        list: List of AudioSegment chunks.
    """
    audio = AudioSegment.from_file(audio_file_path)
    chunks = split_on_silence(
        audio, 
        min_silence_len=min_silence_len,
        silence_thresh=silence_thresh,
        keep_silence=keep_silence
    )
    return chunks

def transcribe(audio_file):
    """
    Transcribe an audio file using the locally loaded Whisper model from Hugging Face.
    
    Args:
        audio_file (str): Path to the audio file.
    
    Returns:
        str: Transcribed text.
    """
    # Load audio using librosa, resampling to 16000 Hz as required by Whisper
    speech, sr = librosa.load(audio_file, sr=16000)
    input_features = processor(speech, sampling_rate=16000, return_tensors="pt").input_features
    # Generate transcription
    predicted_ids = model.generate(input_features)
    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
    return transcription

def process_audio_chunks(audio_chunks):
    """
    Process and transcribe each audio chunk.
    
    Args:
        audio_chunks (list): List of AudioSegment chunks.
    
    Returns:
        str: Combined transcription from all chunks.
    """
    transcriptions = []
    min_length_ms = 100  # Minimum length required (0.1 seconds)
    
    for i, chunk in enumerate(audio_chunks):
        if len(chunk) < min_length_ms:
            st.warning(f"Chunk {i} is too short to be processed.")
            continue
        
        with NamedTemporaryFile(delete=False, suffix=".wav") as temp_audio_file:
            chunk.export(temp_audio_file.name, format="wav")
            temp_audio_file_path = temp_audio_file.name
        
        transcription = transcribe(temp_audio_file_path)
        if transcription:
            transcriptions.append(transcription)
            st.write(f"Transcription for chunk {i}: {transcription}")
        
        os.remove(temp_audio_file_path)
    return " ".join(transcriptions)

def save_transcription_to_docx(transcription, audio_file_path):
    """
    Save the transcription as a .docx file.
    
    Args:
        transcription (str): Transcribed text.
        audio_file_path (str): Path to the original audio file for naming purposes.
    
    Returns:
        str: Path to the saved .docx file.
    """
    base_name = os.path.splitext(os.path.basename(audio_file_path))[0]
    output_file_name = f"{base_name}_full_transcription.docx"
    doc = Document()
    doc.add_paragraph(transcription)
    doc.save(output_file_name)
    return output_file_name

st.title("Audio Transcription with Whisper (Local via Hugging Face)")

# Allow uploading of audio or video files
uploaded_file = st.file_uploader("Upload an audio or video file", type=["wav", "mp3", "ogg", "m4a", "mp4", "mov"])

if 'transcription' not in st.session_state:
    st.session_state.transcription = None

if uploaded_file is not None and st.session_state.transcription is None:
    st.audio(uploaded_file)
    
    # Save uploaded file temporarily
    file_extension = uploaded_file.name.split(".")[-1]
    temp_audio_file = f"temp_audio_file.{file_extension}"
    with open(temp_audio_file, "wb") as f:
        f.write(uploaded_file.getbuffer())
    
    # Split and process audio using silence detection
    with st.spinner('Transcribing...'):
        audio_chunks = split_audio_on_silence(
            temp_audio_file, 
            min_silence_len=500,
            silence_thresh=-40,
            keep_silence=250
        )
        transcription = process_audio_chunks(audio_chunks)
        if transcription:
            st.session_state.transcription = transcription
            st.success('Transcription complete!')
            output_docx_file = save_transcription_to_docx(transcription, uploaded_file.name)
            st.session_state.output_docx_file = output_docx_file
    
    if os.path.exists(temp_audio_file):
        os.remove(temp_audio_file)

if st.session_state.transcription:
    st.text_area("Transcription", st.session_state.transcription, key="transcription_area_final")
    with open(st.session_state.output_docx_file, "rb") as docx_file:
        st.download_button(
            label="Download Transcription (.docx)",
            data=docx_file,
            file_name=st.session_state.output_docx_file,
            mime='application/vnd.openxmlformats-officedocument.wordprocessingml.document'
        )