Seshat TTS

Seshat TTS is a Windows GUI utility for realtime audio streaming for games, or apps. Pick a monitor or window, drag one capture region over the text, press one hotkey, and the selected text is extracted with Tesseract OCR or a local vision LLM, then streamed through Kyutai Pocket TTS.

Maintained by Scriptriva Inc.

For support inquiries email: support@scriptriva.com

Demo

What It Does

• Captures one selected screen region from a monitor or a chosen window.
• Runs Tesseract OCR on that exact region, or sends the region image directly to a local vision-capable LLM for text extraction.
• Streams the extracted text through Pocket TTS in realtime.
• Lets you use a built-in Pocket TTS voice for speed or upload a custom WAV/MP3 reference voice.
• Optionally routes OCR text through a local OpenAI-compatible LLM endpoint before speech.
• Includes a 0-300% playback volume slider for quiet voices or noisy games.
• Stops any active audio stream when a new read starts, so repeated hotkey presses do not overlap.
• Caches custom voice state as .safetensors for faster repeat custom-voice reads when using the uvx-server backend.

Requirements

• Windows 10/11.
• Python 3.10 through 3.14 when running from source or building.
• Tesseract OCR for Windows when running from source or building a portable EXE with bundled OCR.
• uvx when running from source, or when building a portable EXE with bundled uvx.
• A working audio output device.

Install Tesseract:

winget install UB-Mannheim.TesseractOCR

Install uvx:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Install Seshat TTS for development or for the fast launcher:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e .[test]

Build Before Use

For a single-file portable EXE, build with:

.\scripts\build_exe.ps1

Portable output:

.\dist\seshat-tts.exe

That EXE bundles the Seshat GUI/runtime files, app resources, uvx.exe if it is available on the build machine, and Tesseract OCR files if Tesseract is installed at C:\Program Files\Tesseract-OCR. You can override the OCR bundle source before building:

$env:SESHAT_TESSERACT_DIR='D:\Tools\Tesseract-OCR'
.\scripts\build_exe.ps1

For the old one-folder PyInstaller build:

.\scripts\build_exe.ps1 -OneDir

One-folder output:

dist\seshat-tts\seshat-tts.exe

The portable EXE still uses Pocket TTS through uvx-server. It does not freeze Torch/Pocket TTS inside the EXE because that path has been unreliable on Windows and can trigger native DLL initialization failures. First Pocket TTS use can still download/cache the Pocket TTS tool and model data under the user's normal cache directories, but no separate Python, Tesseract, or uvx install should be needed when those files were bundled during build.

For a tiny development launcher, build:

.\scripts\build_launcher_exe.ps1

Launcher output:

dist\launcher\seshat-tts.exe

This launcher is intentionally small and quick to build. It uses the .venv in this project when present, so keep the virtual environment and installed dependencies beside the launcher.

Run From Source

seshat-tts

For the fast launcher EXE, run:

.\dist\launcher\seshat-tts.exe

The launcher expects dependencies in .venv or your active Python environment. It does not bundle Python, Torch, Pocket TTS, or Tesseract.

First-Time Setup

1. Open Seshat TTS.
2. Choose monitor or window capture mode.
3. Select the monitor or window to watch.
4. Click Select Region, then drag over the exact text area to read.
5. Click inside Read Hotkey and press the key combo you want. The default is ctrl+alt+n.
6. Click inside Region Hotkey and press the key combo you want. The default is ctrl+alt+r.
7. Click inside Stop Hotkey and press the key combo you want. The default is ctrl+alt+s.
8. Set Tesseract if it was not detected automatically.
9. Choose a voice:
   • default is fastest and uses a built-in Pocket TTS voice.
   • custom-wav lets you choose a named WAV, MP3, or cached .safetensors reference voice.
10. Adjust Volume if the generated voice is too quiet. 100% is neutral; values above that boost and clip safely.
11. Enable Local LLM if you want OCR text cleaned by a local OpenAI-compatible server before TTS.
12. Enable Use local LLM vision instead of Tesseract OCR only when your local model endpoint supports image input and you want the LLM to read the selected region directly.
13. Click Preload TTS once before playing if you want the first read to be less delayed.
14. Press the read hotkey whenever the selected text should be spoken, or the stop hotkey whenever playback should stop.

Use borderless/windowed mode for games if exclusive fullscreen capture returns stale or blank frames.

Local LLM

The Local LLM panel can use an OpenAI-compatible endpoint in two ways:

• Route OCR through local OpenAI-compatible LLM keeps Tesseract as the text extractor, then asks the local model to clean the parsed text before TTS.
• Use local LLM vision instead of Tesseract OCR skips Tesseract and sends the selected region image to the local model as a PNG data URL. This requires a vision-capable OpenAI-compatible model endpoint.

Typical values:

Base URL: http://127.0.0.1:8000/v1
API Key: local key or token
Model: the model name exposed by your local server

Load api_key.txt fills the API key field from a repo-local api_key.txt file if present. Treat that file as a secret and do not commit it. Lower timeout and max token values reduce latency; no network or LLM path can be truly zero-latency, but a local endpoint keeps this as short as the model server allows.

Disable thinking is enabled by default. It sends common OpenAI-compatible metadata for local reasoning models, including chat_template_kwargs.enable_thinking=false, so models that support that switch skip reasoning output and return faster.

Voice Modes

default voice mode is the fastest. Pick a built-in voice such as alba, marius, anna, vera, or george.

custom-wav mode accepts .wav, .mp3, and cached .safetensors voice files. MP3 references are converted once into cached WAV files before Pocket TTS processes them. Use Manage beside Custom Voice to name voices, save them, and select them from the dropdown.

The first custom-voice run can be slow because Pocket TTS must convert the reference audio into a voice state. Seshat TTS caches that state under:

%USERPROFILE%\.seshat-tts\voices

After that cache exists, the uvx-server backend sends a reusable local voice_url instead of uploading and reprocessing the same audio every time. Named custom voices are stored in:

%USERPROFILE%\.seshat-tts\voice_profiles.json

Pocket TTS voice cloning may require Hugging Face access:

1. Request access on Kyutai's Pocket TTS Hugging Face page.
2. Create a token at Hugging Face tokens.
3. Login for uvx:

uvx hf auth login --force

Build Commands

Fast launcher build, usually under a minute:

.\scripts\build_launcher_exe.ps1

Output:

dist\launcher\seshat-tts.exe

Full dependency-bundled PyInstaller build:

.\scripts\build_exe.ps1

Output:

dist\seshat-tts.exe

Use the fast launcher during development and for local use. Use the portable build when you need to move the app to a machine where Python, Tesseract, and uvx are not installed.

The python-api backend is only shown when running from source or the fast launcher. The bundled PyInstaller EXE only exposes uvx-server.

License and Reuse

Seshat TTS is released under the Scriptriva Public Source License 1.0.

Copyright (c) 2026 Scriptriva Inc.

Commercial use is allowed under the license terms. The license preserves attribution, third-party notices, Scriptriva branding rights, safety restrictions, and restrictions on reusing the licensed work to create or distribute a same-functionality product.

Useful reuse boundaries:

• src/seshat_tts/capture.py: monitor/window capture helpers.
• src/seshat_tts/ocr.py: OCR preprocessing and text extraction.
• src/seshat_tts/tts.py: Pocket TTS server/API playback adapters and stream cancellation.
• src/seshat_tts/llm.py: OpenAI-compatible local LLM cleanup step.
• src/seshat_tts/config.py: persisted GUI/runtime configuration.
• src/seshat_tts/region_picker.py: snipping-tool-style region selection.

Security and privacy considerations for reuse:

• Treat OCR text, API keys, custom voice files, and generated voice caches as user data.
• Do not commit api_key.txt, voice samples, .safetensors voice caches, or local config files.
• Custom voice cloning should be used only with audio you have permission to use.
• The portable EXE may bundle third-party binaries; keep their notices and license terms intact.

Third-Party Notices

Seshat TTS uses and/or interfaces with these third-party projects. Each project remains under its own license:

Component	Purpose	License	Notes
Kyutai Pocket TTS	Local text-to-speech generation and voice cloning	MIT	The Pocket TTS GitHub repository identifies the project as MIT licensed. Model/voice assets may have separate terms; review the linked Hugging Face pages before redistribution.
Tesseract OCR	OCR engine used to extract text from selected screen regions	Apache License 2.0	Tesseract is not MIT licensed. Its project site identifies it as Apache 2.0 licensed.
pytesseract	Python wrapper for Tesseract	Apache License 2.0	Used to invoke the Tesseract executable from Python.
PyInstaller	Windows executable packaging	GPLv2-or-later with bootloader exception	Used only for building packaged executables.
OpenAI Python SDK	OpenAI-compatible local LLM client	Apache License 2.0	Used for optional local LLM cleanup through OpenAI-compatible endpoints.

Packaged builds include THIRD_PARTY_NOTICES.md, including a link to the Pocket TTS MIT license.

Tests

$env:PYTHONPATH='src'
python -m pytest -q