services/whisper/utils/broll.py

"""
broll.py
---------------------------------------
AI B-Roll Injection System (V8)

Purpose:
- Detect topics in transcript
- Map topics → generic stock B-roll assets
- Overlay or replace segments
- Improve retention & visual variety

Works in CPU-only environments.
No external API dependency required.
"""

import os
import random
import subprocess


# =====================================================
# STOCK B-ROLL LIBRARY (LOCAL FALLBACK)
# =====================================================

DEFAULT_BROLL = {
    "money": "assets/broll/money.mp4",
    "success": "assets/broll/success.mp4",
    "business": "assets/broll/business.mp4",
    "phone": "assets/broll/phone.mp4",
    "tech": "assets/broll/tech.mp4",
    "people": "assets/broll/people.mp4",
    "talking": "assets/broll/talking.mp4",
    "default": "assets/broll/default.mp4",
}


# =====================================================
# TOPIC DETECTION
# =====================================================

def detect_topic(text):
    """
    Simple keyword-based topic classifier.
    Lightweight (no ML dependency).
    """

    text = text.lower()

    if any(w in text for w in ["money", "rich", "income", "profit"]):
        return "money"

    if any(w in text for w in ["business", "startup", "company"]):
        return "business"

    if any(w in text for w in ["phone", "mobile", "iphone", "android"]):
        return "phone"

    if any(w in text for w in ["tech", "ai", "software", "computer"]):
        return "tech"

    if any(w in text for w in ["success", "win", "achieve"]):
        return "success"

    if any(w in text for w in ["people", "person", "man", "woman"]):
        return "people"

    if any(w in text for w in ["talk", "speak", "say"]):
        return "talking"

    return "default"


# =====================================================
# SEGMENT ANALYZER
# =====================================================

def extract_segments(words, segment_length=8):
    """
    Converts transcript words into grouped segments.
    """

    segments = []
    buffer = []

    for w in words:
        buffer.append(w)

        if len(buffer) >= segment_length:
            segments.append(buffer)
            buffer = []

    if buffer:
        segments.append(buffer)

    return segments


# =====================================================
# B-ROLL MATCHING ENGINE
# =====================================================

def match_broll(segment):
    """
    Map transcript segment → B-roll video
    """

    text = " ".join([w["word"] for w in segment])
    topic = detect_topic(text)

    return DEFAULT_BROLL.get(topic, DEFAULT_BROLL["default"])


# =====================================================
# B-ROLL INSERTION (FFMPEG OVERLAY STRATEGY)
# =====================================================

def overlay_broll(base_video, broll_video, output_path, start_time, duration):
    """
    Overlays B-roll using ffmpeg.
    Lightweight crossfade approach.
    """

    cmd = [
        "ffmpeg", "-y",
        "-i", base_video,
        "-i", broll_video,
        "-filter_complex",
        f"[1:v]scale=1080:1920,format=rgba[ov];"
        f"[0:v][ov]overlay=enable='between(t,{start_time},{start_time+duration})'",
        "-c:v", "libx264",
        "-preset", "ultrafast",
        "-c:a", "copy",
        output_path
    ]

    subprocess.run(cmd, check=True)


# =====================================================
# MAIN PIPELINE
# =====================================================

def insert_broll(video_path, words=None):
    """
    Full B-roll injection pipeline
    """

    if not words:
        # fallback: return original video
        return video_path

    segments = extract_segments(words)

    current_video = video_path
    outputs = []

    for i, segment in enumerate(segments):

        broll = match_broll(segment)

        output_file = f"broll_output_{i}.mp4"

        start_time = segment[0]["start"]
        duration = segment[-1]["end"] - start_time

        try:
            overlay_broll(
                current_video,
                broll,
                output_file,
                start_time,
                duration
            )

            current_video = output_file
            outputs.append(output_file)

        except Exception as e:
            print(f"[BROLL ERROR] Segment {i}: {e}")
            continue

    return outputs[-1] if outputs else video_path


# =====================================================
# ADVANCED VERSION (V8 EXTENSION)
# =====================================================

def smart_broll_engine(words, hook_boost=True):
    """
    Enhanced version:
    - prioritizes hook segments
    - increases emotional pacing
    """

    segments = extract_segments(words)

    prioritized = []

    for seg in segments:

        text = " ".join([w["word"] for w in seg]).lower()

        score = 0

        if any(k in text for k in ["you", "this", "stop", "now"]):
            score += 2

        if hook_boost and len(seg) < 5:
            score += 1

        prioritized.append((score, seg))

    prioritized.sort(reverse=True, key=lambda x: x[0])

    final_video = None

    for _, seg in prioritized:
        final_video = insert_broll(final_video or "input.mp4", seg)

    return final_video