PDF to Audio 0.2.0

PDF to Audio 0.2.0

Convert PDFs and text documents into local, voice-cloned narration.

• Six-excerpt cleanup and six-passage speech batches on CPU and GPU; GPUs with 4 GiB of VRAM or less use one at a time.
• Both models can use all CPU threads, including during GPU inference.
• Cleanup and speech models run in separate stages and unload between them; audio decoding is sequential to reduce peak VRAM.
• Elapsed conversion timer, saved queues, and resumable cleanup/audio generation.
• Original-language narration for mixed-language document batches, with scientific cleanup instructions for acronyms, equations, and figure/table captions.
• Every file goes through Qwen3 cleanup before narration; PDFs first go through text extraction. Table contents are omitted while captions are preserved.
• First-launch setup shows download percentage and transferred bytes, with resumable runtime downloads.
• Default 0.6B models download during first-launch setup; optional 1.7B models download from Settings.

Install

• Ubuntu 24.04 or newer: open the .deb with your software installer.
• Fedora 44: open the .rpm with your software installer.
• Windows x64: run PDF-to-Audio-Windows-x64-Setup.exe.
• macOS Apple Silicon: open the .dmg and drag PDF to Audio into Applications.

Then open PDF to Audio and let its setup download its private runtime and the two default models (about 4 GB for models, plus runtime dependencies). No Python installation or terminal commands are required. Internet is required for setup, not narration.

The same Linux/Windows installer supports CPU and NVIDIA CUDA 12.8. Setup detects an available NVIDIA driver and downloads CUDA dependencies; choosing CUDA in Settings also installs them if missing. Installing a driver later does not require reinstalling the app. CPU mode works without GPU drivers. A compatible host NVIDIA driver is required for GPU acceleration and is not installed by the app.

The Apple Silicon installer includes CPU and experimental Apple Metal (PyTorch MPS) support. Automatic mode selects Metal on supported physical GPUs, using float32 and CPU fallback for unsupported operators. Both models release their GPU caches between stages. CPU narration is tested; Metal hardware inference is not yet validated on a physical Mac. Qwen failed on the hosted runner's Apple Paravirtual GPU, so virtual GPUs now use CPU and explicit Metal selection reports a clear error. AMD and Intel GPU acceleration on Linux/Windows is not supported.

Windows/macOS files are unsigned and macOS is not notarized, so the OS may display a security warning or require explicit approval through its security settings. Do not disable system security globally.

Release validation: Linux and Windows installers are tested automatically. Final macOS installer validation is deferred to manual testing; the preceding macOS build passed CPU narration, but failed Metal inference on the virtual GPU.

Scanned PDFs need OCR first. Review important narration for extraction/cleanup errors. Resume requires an unchanged document, voice, language, and cleanup prompt. Models can change: completed cleanup and audio are kept, while new models handle unfinished work. An interrupted batch is regenerated; switching speech models may change the sound of later passages.