AutoTalk

Turn PDF slides into a timed, spoken presentation.

AutoTalk is a Python/PySide6 desktop application. Codex app-server writes narration using your ChatGPT sign-in; Qwen3-TTS 1.7B generates speech on your local GPU. The executable is autotalk (autotalk.exe on Windows).

The expanded 0.2 prototype implements the agreed feature list. Read verification results and platform limitations before relying on it for a conference.

Start

Extract the native package and open autotalk, autotalk.exe, or AutoTalk.app. Keep the accompanying files together. In this checkout, ./autotalk launches the local Linux bundle at dist/autotalk/autotalk.

Platform	Speech backend	Qualification
Linux x86_64, NVIDIA GPU	vLLM-Omni streaming	Tested locally on an RTX 2000 Ada laptop, 8 GB VRAM
Windows 11 x64, NVIDIA GPU	Qwen/PyTorch CUDA, sentence chunks	Implemented; native build and GPU testing pending
macOS 14+, Apple Silicon	MLX Audio, Metal streaming	Implemented; native build and GPU testing pending

The application packages Python, Qt, and audio/video libraries. Speech setup privately downloads its Python runtime, pinned dependencies, models, and any required compiler. Users do not install Python, a CUDA toolkit, or speech servers. A supported GPU with a working compatible OS driver is required for synthesis. CPU/NPU inference and Intel Macs are outside this prototype. Prepared playback does not require the speech runtime or a GPU capable of synthesis.

First setup requires internet access and several GB of downloads. Allow roughly 40 GB of free space for the Linux runtime, model, and installation caches; more for additional models and projects. Downloads are cached and verified. Cancelling retains completed slide audio; an interrupted slide is regenerated on retry. Runtime/model data stays in the user's application-data directory.

Three modes

1. Prepared: import a PDF, choose language, duration, conference context and voice, create and review narration, then generate audio. Optionally fit its measured duration with up to three revision passes. Start fullscreen explicitly.
2. Quick: choose the PDF, language, and duration, then Start. AutoTalk uses Ryan, professional delivery, default Codex settings, and no conference context. The default policy attempts duration fitting, then presents even if a visible mismatch remains. Alternatives generate once or require a timing match.
3. Realtime: prepare the opening audio buffer and start presenting while later audio is generated. Choose a complete script first, or a whole-deck outline followed by two-slide writing batches overlapping synthesis. Buffer underruns wait visibly. Final duration remains approximate until preparation finishes.

First-time downloads and model loading precede speech in every mode. Streaming reduces waiting for audio; it does not guarantee uninterrupted playback on every GPU.

Configure the talk

• Import an unencrypted PDF of 1–80 slides into a new project directory.
• Select duration in minutes, tolerance, inter-slide pause, audience, and objective.
• Enter conference scope directly, read it from a URL, or combine both. Review the editable extraction and its source links. JavaScript-only and authenticated websites may need a manually supplied explanation.
• Use Connect ChatGPT for subscription sign-in and model discovery. The model selector lists account-available models accepting slide images; reasoning choices come from the selected model. Subscription limits apply. No paid API fallback is used. Slide images, extracted text, and supplied context are sent to Codex; reference recordings remain local.
• Select English, German, French, Spanish, Italian, Portuguese, Russian, Chinese, Japanese, or Korean. Add language versions to retain independent scripts/audio. [German] and similar paragraph markers allow mixed passages. Choose mixed passages, one language per slide, or a single language per version in settings.

Voices and delivery

Predefined voices use CustomVoice: Ryan, Aiden, Vivian, Serena, Uncle_Fu, Dylan, Eric, Ono_Anna, and Sohee. Preview in the selected talk language.

My voice uses Base with a clean 3–60 second WAV recording. Record directly (up to 30 seconds), or import a recording. Its exact transcript is recommended; without one, conditioning uses speaker identity alone. No fine-tuning is needed. Save named voices for reuse; project copies retain their own references, organized by language, with a shared reference available for cross-language use.

Design a voice uses VoiceDesign for a preview/reference. Save an accepted voice or prepare the talk to freeze one reference and reuse it through Base. This avoids independently redesigning the speaker on every slide. Base inherits the reference's identity/delivery and does not accept CustomVoice/VoiceDesign vocal instructions.

Choose Professional, Conversational, Energetic, Calm and understated, Lightly humorous, Academic, Storytelling, or Inspirational. Customize acoustic attributes, voice-design age, persona, gradual progression, and per-slide directions. Save named delivery presets. Unsupported controls are disabled. Dialect and expressive cues are model requests requiring preview, not guaranteed effects or sound tags. Use imported clips for precisely controlled singing, music, or nonverbal effects.

Advanced settings expose sampling temperature, top-k, top-p, repetition penalty, and an output token limit. They are not reasoning efforts or guaranteed quality levels. Synthesis stays at native 24 kHz; export resampling and AAC bitrate are separate compatibility/encoding choices.

Present, route audio, and record

Choose the fullscreen display. Use System audio settings and Test audio to select AutoTalk's output through Plasma, Windows Sound settings, or macOS Sound. The operating system owns routing; projects do not store audio-device bindings.

Space pauses/resumes; arrows navigate. Esc pauses and leaves fullscreen, keeping the slide and position. Use Continue presentation, Restart from beginning, or Stop. Previews, narration, imported clips, and background tracks share one PCM transport. Slide advancement follows consumed audio rather than word estimates.

Attach clips before/after slides and adjust their gain. Add a background track, gain, and looping. Fixed clips and pauses count toward duration fitting.

Enable Record presentation as a video under Delivery & recording. Export uses AutoTalk's own audio mix and slide timeline, without recording other desktop sound. MP4 output is 1080p/30 fps with AAC at 24, 44.1, or 48 kHz. Recording policies retain fullscreen pauses only (default), all elapsed waiting, or content only. Configured filenames receive a session suffix to preserve earlier recordings. Export runs after playback; Export recorded session can recover a saved session after a cancelled export or interrupted application session.

Projects and persistence

Save with Ctrl+S and copy the entire project folder to move a talk:

• talk.autotalk.json: versioned configuration, language versions, and artifact metadata.
• slides.pdf, slides/: source PDF and rendered previews.
• voice/<language>/: reference snapshots.
• audio/<version>/<language>/: generated speech, keyed by effective synthesis inputs.
• media/: normalized imported clips/background tracks.
• recordings/<session>/: recoverable PCM, slide images, timeline, and exported video.

Version-1 projects migrate on save with an untouched .v1-backup.json. Verified legacy audio remains playable with its original model provenance. Regeneration uses 1.7B. Changed inputs invalidate affected audio; damaged files are detected on reopening. Old artifacts are retained; automatic disk cleanup is not implemented.

Development and native builds

python3 -m venv .venv
.venv/bin/python -m pip install -e '.[dev]'
QT_QPA_PLATFORM=offscreen .venv/bin/python -m pytest -q
.venv/bin/python tools/build.py

On Windows use .venv\Scripts\python.exe; build on each target OS. The native CI matrix builds Linux archives, Windows ZIP packages, and Apple Silicon DMGs. Tests marked audio_device need a working physical/virtual audio output and run locally; remaining tests run in the native CI matrix. Signing hooks accept AUTOTALK_SIGN_IDENTITY / AUTOTALK_NOTARY_PROFILE on macOS and AUTOTALK_SIGN_CERT / AUTOTALK_TIMESTAMP_URL on Windows.

project.py owns configuration and artifact validity. services.py operates on job snapshots and publishes validated results to Qt. runtime.py owns managed processes; speech_worker.py adapts the three native synthesis backends. playback.py owns presentation position and consumed PCM; media.py records and exports that same stream. The old separate media-player path has been removed.

See third-party notices. This is a development build; public release licensing, native signing, and broader hardware qualification remain open.