whiskeycoder/qwen3-audiobook-converter

Purpose and Entry Point

This is a local Python batch converter that turns documents into MP3 audiobooks through a locally running Qwen3 TTS Gradio service. The CLI entry point is audiobook_converter.py; it processes all supported source files placed under book_to_convert/ and writes finished audio to audiobooks/.

Paths: audiobook_converter.py, book_to_convert/input_here.txt, audiobooks/output_here.txt

Runtime and Dependencies

The application requires Python 3.8+, FFmpeg for audio encoding/assembly, and a Qwen Voice/Gradio server reachable locally at http://127.0.0.1:7860. Python dependencies support Gradio API calls, document parsing, and audio composition.

Paths: requirements.txt, config.py, audiobook_converter.py

CLI Mode Selection and Configuration

The converter operates in default custom-speaker narration mode or optional voice-clone mode, enabled with --voice-clone and a reference WAV via --voice-sample. Configuration centralizes Qwen endpoint/model and fixed audiobook-oriented settings such as speaker, language, chunk size, retries, sequential processing, MP3 format, and bitrate.

Paths: audiobook_converter.py, config.py

Input Discovery and Text Extraction

The batch pipeline discovers files in the input directory and extracts readable text based on extension. Supported boundaries are TXT, PDF, EPUB, DOCX, and legacy DOC; image-only or corrupted documents are expected to fail extraction rather than receive OCR. Extraction dependencies include PyPDF2, EbookLib/BeautifulSoup, python-docx, and docx2txt.

Paths: audiobook_converter.py, book_to_convert/input_here.txt, requirements.txt

Chunking and Synthesis Boundary

Extracted text is split into roughly 1,200-word units while preserving sentence boundaries. Chunks are processed sequentially (MAX_WORKERS=1) to avoid overloading or rate-limiting the local model. Each chunk crosses the principal external boundary: a Gradio client request to Qwen’s 1.7B TTS model; clone mode additionally relies on Qwen transcription of the reference audio.

Paths: audiobook_converter.py, config.py

Audio Assembly, Caching, and Recovery

Generated chunk audio is cached to avoid repeating identical synthesis work, retried on transient failures, then concatenated and encoded as a 128 kbps MP3 through pydub/FFmpeg. The final artifact is written per input book; intermediate chunk files are intended to be cleaned up even if a conversion fails.

Paths: audiobook_converter.py, audiobooks/output_here.txt, requirements.txt

Operational Outputs

The repository ships placeholder input/output directories and a sample text-to-MP3 artifact for validation. Runtime-created cache, chunk, and log directories are described by the project but are not present in this tree, so they are ephemeral operational state rather than versioned application components.

Paths: sample/test_audio.txt, sample/test_audio.mp3, README.md