AutoTalk v0.1

AutoTalk’s first public release turns PDF slide decks into timed, spoken presentations. Prepare narration with Codex, generate speech locally with Qwen3-TTS 1.7B, and present your slides with synchronized audio in a native Qt desktop application.

Prepare and present your talk

• Prepared: edit and review narration and audio before presenting.
• Quick: choose a PDF, language and duration, then start automatic preparation and presentation.
• Realtime: begin after the opening audio buffer is ready while later slides continue preparing.

Tailor the talk to your audience using a conference website or written context. Create and edit narration per slide or for the whole deck, preview individual slide audio, and retain separate language versions. Choose a predefined voice, use your own reference recording, or design a voice.

AutoTalk includes fullscreen presentation controls, pause/resume and live-demo breaks, reusable saved projects and audio, video/audio export, and Linux desktop recording. Speech-model loading and unloading are controllable from the application.

Downloads and getting started

Download the package for your platform below:

Platform	Package	Speech generation
Linux x86_64	.tar.gz	NVIDIA GPU; vLLM-Omni
Windows x64	.zip	NVIDIA GPU; Qwen / PyTorch CUDA
macOS Apple Silicon	.dmg	MLX Audio / Metal

Extract the Linux or Windows archive and launch autotalk or autotalk.exe; on macOS, open the disk image and copy AutoTalk to Applications. SHA-256 checksum files accompany each package. Package metadata uses version 0.1.0 for release v0.1. These CI packages do not include publisher signing or Apple notarization.

Open a PDF, select the talk language and target duration, choose a voice and presentation mode, and sign in to Codex when prompted. Initial speech setup downloads the runtime and models automatically and requires internet access, compatible GPU drivers, and several GB of downloads. Allow approximately 40 GB of free space on Linux for the runtime, models and installation caches. Prepared audio can play without loading a speech model.

Release status

This is an initial release. Linux, Windows and macOS pass automated tests, package builds and packaged-application startup checks. Linux has also been exercised on real hardware; end-to-end Windows/macOS speech, GPU, audio-device and recording workflows still require native hardware qualification. Realtime startup depends on model loading and synthesis speed; the requested talk duration is a target rather than an exact guarantee.

Narration generation uses Codex app-server with an eligible ChatGPT sign-in and is subject to account usage limits. Slide images, extracted text and supplied context are sent to Codex; voice references and speech synthesis remain local.

User guide · Build guide · Source and screenshots

Released under the MIT License. Third-party dependencies retain their own licenses; see the included THIRD_PARTY.md.

Release packages are built and smoke-tested in the v0.1 CI run.