scarif/pdf-to-audiobook/pdf_to_audiobook.py

#!/usr/bin/env python3
"""
pdf_to_audiobook.py

1) Always convert PDF → Markdown via marker + your OpenAI-compatible endpoint.
2) If you pass --voice-name, use Coqui XTTS (voice cloning);
   otherwise use Kokoro TTS with the “Daniel” speaker.
3) Manage your 6-sec snippets in voices/ directory.
"""

import os
import argparse
import tempfile
import numpy as np
import torch
import soundfile as sf
from pathlib import Path
from pdfminer.high_level import extract_text
from pydub import AudioSegment# Allow Coqui’s XttsConfig class to be unpickled

# ─── Whitelist Coqui classes for safe CPU‐only loading ────────────────────────
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import XttsAudioConfig
from TTS.config.shared_configs import BaseDatasetConfig
from TTS.tts.models.xtts import XttsArgs
from torch.serialization import add_safe_globals
add_safe_globals([XttsConfig, XttsAudioConfig, BaseDatasetConfig, XttsArgs])

# ─── Coqui XTTS API ─────────────────────────────────────────────────────────
from TTS.api import TTS

# ─── Kokoro TTS API ─────────────────────────────────────────────────────────
from kokoro import KPipeline

# ─── marker PDF→Markdown API ────────────────────────────────────────────────
from marker.converters.pdf import PdfConverter
from marker.models           import create_model_dict
from marker.config.parser    import ConfigParser
from marker.output           import text_from_rendered

VOICE_DIR = Path("voices")

def list_available_voices():
    return sorted(p.stem for p in VOICE_DIR.glob("*.wav"))

def convert_pdf_to_markdown(pdf_path: str) -> str:
    key = os.getenv("OPENAI_API_KEY")
    model = os.getenv("OPENAI_MODEL")
    url = os.getenv("OPENAI_BASE_URL")
    if not key:
        raise RuntimeError("Please set OPENAI_API_KEY in your environment")
    cfg = {
        "output_format":   "markdown",
        "use_llm":         True,
        "llm_service":     "marker.services.openai.OpenAIService",
        "openai_api_key":  key,
        "openai_model":    model,
        "openai_base_url": url,
    }
    parser    = ConfigParser(cfg)
    converter = PdfConverter(
        config         = parser.generate_config_dict(),
        artifact_dict  = create_model_dict(),
        processor_list = parser.get_processors(),
        renderer       = parser.get_renderer(),
        llm_service    = parser.get_llm_service(),
    )
    rendered = converter(pdf_path)
    markdown, _, _ = text_from_rendered(rendered)
    return markdown

def synthesize_with_coqui(
    text: str,
    out_mp3: str,
    voice: str,
    speaker: str | None,
    lang:    str,
):
    use_gpu = torch.cuda.is_available()
    tts     = TTS(model_name="tts_models/multilingual/xtts_v2", gpu=use_gpu)

    spk_wav  = [voice]
    spk_name = speaker or tts.speakers[0]
    print(f"⚙️ Using Coqui voice sample “{voice}” → speaker “{spk_name}”")

    tmp_wav = tempfile.mktemp(suffix=".wav")
    tts.tts_to_file(
        text            = text,
        file_path       = tmp_wav,
        speaker_wav     = spk_wav,
        speaker         = spk_name,
        language        = lang,
        split_sentences = True,
    )
    AudioSegment.from_wav(tmp_wav).export(out_mp3, format="mp3")
    print(f"✅ Coqui MP3 saved → {out_mp3}")

def synthesize_with_kokoro(
    text:   str,
    out_mp3:str,
    lang:   str,
):
    pipeline = KPipeline(lang_code=lang)
    chunks = []
    # Daniel is the chosen voice
    for _, _, audio in pipeline(text, voice="Daniel"):
        chunks.append(audio)
    full = np.concatenate(chunks, axis=0)

    tmp_wav = tempfile.mktemp(suffix=".wav")
    sf.write(tmp_wav, full, 24000)
    AudioSegment.from_wav(tmp_wav).export(out_mp3, format="mp3")
    print(f"✅ Kokoro MP3 saved → {out_mp3}")

def main():
    # ensure voices/ exists
    VOICE_DIR.mkdir(exist_ok=True)

    ap = argparse.ArgumentParser(
        description="PDF→Markdown→Audiobook (Coqui or Kokoro/Daniel)"
    )
    ap.add_argument("pdf",           help="Input PDF file")
    ap.add_argument("--list-voices", action="store_true",
                    help="Show available voice snippets and exit")
    ap.add_argument("--voice-name",  help="Name of a .mp3 in voices/ to clone")
    ap.add_argument("--speaker",     help="Coqui built-in speaker (default if none)")
    ap.add_argument("--lang",        default="en",  help="Language code")
    ap.add_argument("--out",         default="audiobook.mp3",
                    help="Output MP3 path")
    args = ap.parse_args()

    if args.list_voices:
        print("Available voices:")
        for name in list_available_voices():
            print("  ", name)
        return

    # 1) PDF→Markdown
    print("🔄 Converting PDF → Markdown…")
    text = convert_pdf_to_markdown(args.pdf)

    # 2) Choose engine
    if args.voice_name:
        wav_path = VOICE_DIR / f"{args.voice_name}.wav"
        if not wav_path.is_file():
            raise FileNotFoundError(f"voices/{args.voice_name}.wav not found.")
        synthesize_with_coqui(text, args.out, str(wav_path),
                              args.speaker, args.lang)
    else:
        synthesize_with_kokoro(text, args.out, args.lang)

if __name__ == "__main__":
    main()