Metadata-Version: 2.4
Name: voizce
Version: 0.1.0
Summary: Private real-time prompter for video calls: hears the questions, answers from your repo with Claude, invisible to screen sharing.
Author: Rafa
License-Expression: MIT
Project-URL: Homepage, https://github.com/rafavth/Voizce
Keywords: claude,meeting-assistant,speech-to-text,parakeet,windows
Classifier: Operating System :: Microsoft :: Windows :: Windows 10
Classifier: Operating System :: Microsoft :: Windows :: Windows 11
Classifier: Environment :: Win32 (MS Windows)
Classifier: Topic :: Multimedia :: Sound/Audio :: Speech
Classifier: Programming Language :: Python :: 3
Requires-Python: <3.13,>=3.11
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: onnx-asr==0.12.0
Requires-Dist: onnxruntime-directml==1.24.4
Requires-Dist: PyAudioWPatch==0.2.12.9
Requires-Dist: silero-vad-lite==0.4.0
Requires-Dist: claude-agent-sdk==0.2.186
Requires-Dist: anthropic==2.12.0
Requires-Dist: PySide6==6.12.0
Requires-Dist: numpy==2.5.5
Requires-Dist: huggingface_hub==2.4.0
Dynamic: license-file

<div align="center">

# 🎙️ Voizce

**Your private prompter for video calls.**
It hears the questions people ask you, transcribes them on your PC and shows you an answer in under 3 seconds,
in a window nobody sees, even when you share your screen.

![Windows](https://img.shields.io/badge/Windows-10%202004%2B%20%7C%2011-0078D4?logo=windows&logoColor=white)
![Python](https://img.shields.io/badge/Python-3.12-3776AB?logo=python&logoColor=white)
![Claude](https://img.shields.io/badge/Claude-Haiku%205.5-D97757?logo=anthropic&logoColor=white)
![ASR](https://img.shields.io/badge/ASR-Parakeet%20(local)-76B900)
![License](https://img.shields.io/badge/license-MIT-green)

**English** · [Español](https://github.com/rafavth/Voizce/blob/main/README.es.md)

<!-- TODO: demo GIF, recorded with `voizce --demo` -->

</div>

---

Made for **defending your own work**: a thesis defense, a technical demo, a client call about your code.
Voizce reads your repository beforehand, so when someone asks *"why did you pick Cloudflare?"* or
*"how do you protect the WebSocket?"*, the answer, based on your actual code, is already on screen.

## ✨ Features

- 🔊 **Only listens to the others.** It captures what plays through your speakers or headphones (WASAPI loopback) and never touches your microphone.
- 🧠 **Local transcription.** Parakeet TDT 0.6b v3 runs on your CPU, so the audio never leaves your machine.
- ❓ **Spots the questions.** A permissive local filter picks candidates, and Claude ignores whatever isn't meant for you.
- ⚡ **Answers in under 3 s** after the other person stops talking, streamed as it's written. Median in testing: 1.6 s.
- 📚 **Knows your project:** a summary of your repository, your own documents and a few lines about you.
- 🙈 **Invisible to screen sharing and recordings** (`WDA_EXCLUDEFROMCAPTURE`), frameless and always on top.
- 🖱️ **Click to answer.** If a question slipped through, click its line in the transcript.
- 💳 **Two engines:** your Claude Code subscription (default) or an Anthropic API key.

## 🚀 Install

You need **Windows 10 (2004+) or 11** and either [Claude Code](https://claude.com/claude-code) logged in
(`claude`, then `/login`) or an [Anthropic API key](https://console.anthropic.com/).

**Starting from scratch**, paste this into PowerShell. It installs [uv](https://docs.astral.sh/uv/), then
Voizce (uv fetches the right Python by itself), then opens it:

```powershell
irm https://astral.sh/uv/install.ps1 | iex; & "$HOME\.local\bin\uv.exe" tool install voizce; & "$HOME\.local\bin\voizce.exe"
```

**Already have uv?**

```powershell
uv tool install voizce
voizce
```

The first time, a settings window opens:

1. **Pick your project's folder.** Claude Code reads it (read-only) and writes a summary, which takes a few minutes.
2. **Write a few lines about yourself** and the situation.
3. **Click *Create shortcuts*** so next time you can open Voizce from the desktop or the Start menu.

> The first launch also downloads the Parakeet model (~650 MB).

## 🎯 Usage

| Control | What it does |
|---|---|
| **▶ Listen / ⏸ Pause** | Starts or stops capturing audio. Nothing is captured while paused. |
| **📌** | Pins the window to the top center, under the webcam. |
| **⚙** | Settings: repository, extra document, about you, API key. |
| **Click a line** | Claude answers that line. |
| **✕** | Closes everything. |

Each answer shows how long it took since the other person stopped talking. If your repository changes, the
status bar tells you; tick *Create the repository summary now* in ⚙ to update it.

## 🔄 How it works

```mermaid
flowchart LR
    A[🔊 System audio<br/>WASAPI loopback] --> B[Silero VAD<br/>splits phrases]
    B --> C[Parakeet<br/>local transcription]
    C --> D{Filter<br/>question?}
    D -- no --> T[Transcript]
    D -- yes --> E[Claude Haiku<br/>+ repo summary]
    E -- SKIP --> T
    E -- answer --> F[🪟 Window hidden<br/>from capture]
    T -. click .-> E
```

**What reaches Claude:** your documents (repository summary, extra document, about you) and, for each
question, that phrase plus the 10 lines before it. Audio and everything else stay on your PC.

**Languages:** Parakeet transcribes 25 European languages and Claude answers in the language of the
question. The question filter is tuned for Spanish and English; in other languages it relies on question
marks, and you can always click a line. The interface is in English or Spanish, following Windows.

## ⚙️ Advanced settings

Everything lives in `%APPDATA%\Voizce`: `config.toml`, the repository summary (`REPO_DIGEST.md`) and
`voizce.log`. Besides what ⚙ shows, `config.toml` accepts:

| Key | Default | What for |
|---|---|---|
| `silence_ms` | `250` | Silence that closes a phrase. The main latency knob: lower is faster, but cuts more phrases in half (Voizce rejoins them) |
| `model` | `claude-haiku-5-5` | Answer model. Haiku is the fastest |
| `timeout_s` | `8` | Seconds before a question is given up |
| `asr` | `cpu-int8` | `directml` to transcribe on the GPU (it was slower on the AMD GPU we tested) |
| `digest_model` | (Claude Code's) | Model that writes the repository summary |

Command line: `voizce digest` rewrites the summary from the terminal, and `voizce --demo` keeps the window
visible to screen recordings so you can record it.

## 🧑‍💻 Development

```powershell
git clone https://github.com/rafavth/Voizce.git
cd Voizce
uv venv; uv pip install -e .
.\.venv\Scripts\python -m voizce
```

## 🗑️ Uninstall

```powershell
uv tool uninstall voizce
```

Then delete `%APPDATA%\Voizce` and the shortcuts.

## ⚠️ Disclaimer

Voizce doesn't record audio, but it does transcribe other people. Depending on the context and where you
live, you may need to tell the other participants. That's your responsibility.

## 📄 License

[MIT](https://github.com/rafavth/Voizce/blob/main/LICENSE)
