🦾 유튜브 쇼츠메이커 개발기

# 1. GPT로 음성 대본 생성
def generate_voice_script(food):
    prompt = f\”\”\”
    Write a 10-second narration for a YouTube Shorts video about ‘{food}’.
    – Do NOT include speaker names or acting cues.
    – No scene directions or sound/music notes.
    – Structure: recommend a recipe (ingredients/time), nutrition/health summary, one reference from literature/movie/real-life, 1-2 curiosity-provoking or self-questioning lines.
    – 30–40 words, all sentences conversational and naturally flowing.
    Write the entire script in natural Korean.
    \”\”\”
    response = client.chat.completions.create(
        model=\”gpt-4o-mini\”,
        messages=[{\”role\”: \”user\”, \”content\”: prompt}]
    )
    return response.choices[0].message.content.strip()

# 2. Typecast API로 텍스트를 음성으로 변환(TTS)
def generate_audio(text, food):
    os.makedirs(\”segments\”, exist_ok=True) # 음성 저장용 폴더 생성
    clean_text = re.sub(r\”\\s+\”, \” \”, text).strip() # 공백 정리
    payload = {
        \”actor_id\”: \”66d91c60da8dd20be59cd40b\”, # 캐릭터(목소리) ID
        \”text\”: clean_text,
        \”lang\”: \”auto\”,
        \”tempo\”: 1,
        \”volume\”: 100,
        \”pitch\”: 0,
        \”xapi_hd\”: True,
        \”max_seconds\”: 300,
        \”model_version\”: \”latest\”,
        \”xapi_audio_format\”: \”mp3\”,
        \”emotion_tone_preset\”: \”angry-1\”
    }
    resp = requests.post(TYPECAST_URL, headers=TYPECAST_HEADERS, json=payload)
    speak_url = resp.json()[\”result\”][\”speak_url\”]
    print(\”speak_url:\”, speak_url)

    # 음성 파일이 준비될 때까지 최대 20회 폴링
    audio_url, extension = None, None
    for i in range(20):
        poll = requests.get(speak_url, headers=TYPECAST_HEADERS)
        result = poll.json()[\”result\”]
        if result.get(\”status\”) == \”done\” and result.get(\”audio\”) and result[\”audio\”].get(\”url\”):
            audio_url = result[\”audio\”][\”url\”]
            extension = result[\”audio\”][\”extension\”]
            print(f\”[{i+1}회차] 음성 URL:\”, audio_url)
            break
        print(f\”[{i+1}회차] 음성 파일 준비 중… 재시도\”)
        time.sleep(1)
    if not audio_url:
        raise RuntimeError(\”20초 내에 음성 파일 URL을 받지 못했습니다.\”)

    path = f\”segments/{food}_voice.{extension}\”
    audio_res = requests.get(audio_url, headers=TYPECAST_HEADERS)
    with open(path, \”wb\”) as f:
        f.write(audio_res.content)
    print(f\”파일 저장 완료: {path}\”)

    seg = AudioSegment.from_file(path, format=extension)
    duration = len(seg) / 1000.0 # 초 단위
    return path, duration

# 3. DALL·E 이미지 생성 및 저장
def download_image(url, filename):
    os.makedirs(\”images\”, exist_ok=True)
    path = os.path.join(\”images\”, filename)
    img_data = requests.get(url).content
    with open(path, \”wb\”) as f:
        f.write(img_data)
    return path

def generate_images(food, voice_script, num_images=2):
    paths = []
    for i in range(num_images):
        prompt = voice_script.strip()
        print(f\”[{i+1}번째 이미지 프롬프트] {prompt}\”)
        res = client.images.generate(
            model=\”dall-e-3\”,
            prompt=prompt,
            n=1,
            size=\”1024×1792\”
        )
        url = res.data[0].url
        filename = f\”{food}_img_{i}.png\”
        img_path = download_image(url, filename)
        paths.append(img_path)
    if not paths:
        raise ValueError(\”❌ 이미지 생성 실패 (프롬프트 없음)\”)
    return paths

# 4. Whisper로 자막 타임스탬프 추출
def transcribe_audio(audio_path):
    print(\”[Whisper] 음성파일에서 자막 타이밍 추출 중…\”)
    model = whisper.load_model(\”medium\”) # base/small/medium/large 가능
    result = model.transcribe(audio_path)
    subtitles = []
    for segment in result[\”segments\”]:
        subtitles.append({
            \”text\”: segment[\”text\”].strip(),
            \”start\”: segment[\”start\”],
            \”end\”: segment[\”end\”]
        })
    return subtitles

# 5. 자막 텍스트 8자/4단어 이하로 분할 및 타이밍 재할당
def split_whisper_line(text, max_chars=8, max_words=4):
    words = text.split()
    result = []
    chunk = []
    for word in words:
        test_chunk = chunk + [word]
        test_text = \”\”.join(test_chunk)
        if len(test_chunk) < max_words and len(test_text) <= max_chars:
            chunk.append(word)
        else:
            if chunk:
                result.append(\” \”.join(chunk))
            chunk = [word]
    if chunk:
        result.append(\” \”.join(chunk))
    return result

def postprocess_whisper_subtitles(subtitles, max_chars=8, max_words=4):
    new_subs = []
    for sub in subtitles:
        lines = split_whisper_line(sub[\”text\”], max_chars, max_words)
        seg_duration = sub[\”end\”] – sub[\”start\”]
        if len(lines) == 0:
            continue
        per_line = seg_duration / len(lines)
        for i, line in enumerate(lines):
            new_subs.append({
                \”text\”: line,
                \”start\”: sub[\”start\”] + i * per_line,
                \”end\”: sub[\”start\”] + (i + 1) * per_line
            })
    return new_subs

# 6. 자막용 이미지 생성 (중앙 정렬, 흰색 글씨 + 검은색 외곽선)
def create_subtitle_image(text, size=(720, 1280), font_path=FONT_PATH_KO, fontsize=80, output_path=’subtitle_temp.png’):
    W, H = size
    img = Image.new(‘RGBA’, size, (0, 0, 0, 0)) # 투명 배경
    draw = ImageDraw.Draw(img)
    font = ImageFont.truetype(font_path, fontsize)
    clean_text = text.replace(‘\”‘, ”).replace(\”‘\”, \”\”)
    wrapped = clean_text
    text_bbox = draw.textbbox((0, 0), wrapped, font=font)
    text_w = text_bbox[2] – text_bbox[0]
    text_h = text_bbox[3] – text_bbox[1]
    pos = ((W – text_w) // 2, (H – text_h) // 2)
    outline_range = 2
    # 외곽선 효과
    for dx in range(-outline_range, outline_range + 1):
        for dy in range(-outline_range, outline_range + 1):
            if dx != 0 or dy != 0:
                draw.text((pos[0] + dx, pos[1] + dy), wrapped, font=font, fill=\”black\”)
    draw.text(pos, wrapped, font=font, fill=\”white\”)
    img.save(output_path)
    return output_path

# 7. 영상 합성 (이미지+자막+음성 싱크)
def assemble_video_with_whisper_subtitles(image_paths, audio_file, food, subtitles, total_duration):
    W, H = 720, 1280 # 영상 해상도
    per_img = total_duration / len(image_paths) # 이미지별 할당 시간
    img_clips = []
    for idx, img in enumerate(image_paths):
        start = idx * per_img
        duration = per_img if idx < len(image_paths) - 1 else total_duration - start
        img_clips.append(
            ImageClip(img).with_start(start).with_duration(duration).resized((W, H))
        )
    main_clip = concatenate_videoclips(img_clips, method=\”compose\”).with_duration(total_duration)

    subtitle_clips = []
    for i, sub in enumerate(subtitles):
        subtitle_path = create_subtitle_image(sub[\”text\”], size=(W, H), output_path=f\”subtitle_temp_{i}.png\”)
        subclip = ImageClip(subtitle_path)
            .with_start(sub[\”start\”])
            .with_duration(sub[\”end\”] – sub[\”start\”])
        subtitle_clips.append(subclip)

    video = CompositeVideoClip([main_clip] + subtitle_clips)
    audio = AudioFileClip(audio_file)
    final = video.with_audio(audio)
    output_filename = f\”{food}_shorts.mp4\”
    final.write_videofile(output_filename, fps=24)

# 8. 메인 함수
def main(food):
    num_images = int(input(\”생성할 이미지 개수를 입력하세요 (예: 4): \”))

    print(\”[1] 음성 대본 생성 중…\”)
    voice_script = generate_voice_script(food)
    print(\”== Generated Voice Script ==\”)
    print(voice_script)
    print(\”============================\”)

    print(\”[2] 음성 파일 생성 중…\”)
    audio_path, duration = generate_audio(voice_script, food)
    print(f\”[2] 음성 파일 저장된 경로: {audio_path}\”)

    print(\”[3] 이미지 생성 중…\”)
    image_paths = generate_images(food, voice_script, num_images=num_images)

    print(\”[4] Whisper로 자막 타임스탬프 추출 및 8자/4단어 이하 분할…\”)
    raw_subtitles = transcribe_audio(audio_path)
    subtitles = postprocess_whisper_subtitles(raw_subtitles, max_chars=8, max_words=4)

    print(\”[5] 영상 합성 중(정확 싱크 자막)…\”)
    assemble_video_with_whisper_subtitles(image_paths, audio_path, food, subtitles, duration)

    print(f\”[✅ 완료] {food}_shorts.mp4 생성 완료!\”)

if __name__ == \”__main__\”:
    food = input(\”키워드를 입력하세요 (예: banana): \”)
    main(food)

import requests
headers = {
  “Authorization”: “Bearer API_TOKEN”,
  “Content-Type”: “application/json”
}
payload = {
  “actor_id”: “캐릭터ID”,
  “text”: “변환할 텍스트”,
  …
}
resp = requests.post(“https://typecast.ai/api/speak”, headers=headers, json=payload)