<p align="center"> <img src="https://raw.githubusercontent.com/lukaszliniewicz/Pandrator/main/pandrator.png" alt="Pandrator" width="180" /> </p>

Pandrator

Pandrator is a local-first workspace for creating audiobooks, subtitles, and voiceovers. It brings document preparation, transcription, optional AI correction and translation, speech generation, review, and export into one browser interface.

Local speech and transcription models can keep your media on your own computer. Cloud language and speech providers are optional.

Quick start

The current release is Pandrator 0.8.7 with Pandrator Manager 0.9.14. The Manager installs Pandrator, launches it, and lets you add or remove speech components later. Docker and WSL are not required.

Windows

1. Download PandratorManager-0.9.14-windows-x86_64.exe.
2. Run it and choose the parent folder for your installation.
3. Let the Manager install Pandrator, then open the browser interface.
4. Under Providers & services, install only the local models you want.

The Windows executable is not Authenticode-signed, so Windows may show Unknown publisher or a SmartScreen warning. The release page collects the checksums for all downloadable files in one SHA256SUMS file.

Linux

Download the AppImage, make it executable, and run it:

chmod +x PandratorManager-0.9.14-x86_64.AppImage
./PandratorManager-0.9.14-x86_64.AppImage

The first launch asks where to create the managed Pandrator folder. On a headless machine, choose the workspace explicitly:

./PandratorManager-0.9.14-x86_64.AppImage \
  setup --workspace /srv/pandrator --no-open

If AppImage mounting is unavailable, prefix the command with APPIMAGE_EXTRACT_AND_RUN=1.

Updating

Download and run the newer Manager file for your operating system. It reuses the remembered installation when possible; if it asks for a folder, choose the same parent folder as before. Then use Review update in the Manager for Pandrator or any installed engine. Projects, generated media, and other user data are kept separately from replaceable application runtimes.

The Manager keeps the application, model environments, caches, services, and data inside the selected workspace. It does not install operating-system packages. Calibre is optional and is needed only for MOBI conversion.

A good first setup

• For a lightweight ready-made voice, start with Kokoro.
• For voice cloning, start with Qwen3 TTS Base or XTTS v2.
• For transcription and subtitles, install CrispASR.
• Add an LLM provider only if you want AI-assisted correction, translation, research, or speech-text optimization.
• Add RVC only if you want to convert generated speech into a trained target voice.

You can begin with one component and add the others whenever you need them.

What you can make

Audiobooks

• Import TXT, PDF, EPUB, DOCX, MOBI, or pasted text.
• Clean difficult documents with deterministic tools, OCR, and an optional reviewable AI workflow.
• Detect structure and chapters, normalize text, and split it for speech.
• Generate speech in segments, compare takes, edit text, regenerate selected passages, and apply optional RVC conversion.
• Export WAV, MP3, Opus, FLAC, or M4B with chapters, metadata, and cover art.

Subtitles and voiceovers

• Start from SRT subtitles or common audio and video files.
• Transcribe speech with word timestamps and optional diarization.
• Correct, translate, split, merge, and retime subtitles in a side-by-side editor.
• Generate and synchronize dubbed speech without changing the source media.
• Export subtitles alone or create dubbed video with original, mixed, or replacement audio and soft, burned, original-language, translated, or bilingual subtitles.

Voices and providers

• Use local TTS services, OpenAI, Google Gemini, or a compatible custom speech endpoint.
• Connect local OpenAI-compatible LLM servers such as LM Studio or supported cloud providers.
• Record or upload reference voices, keep transcripts with them, and preview built-in voices before generation.
• Import RVC models and retain both original and converted takes.

Choosing local models

The Manager shows the available compute variants, model licences, and downloads before making any changes.

| Component | Best suited to | Typical compute | |---|---|---| | Kokoro 82M | Lightweight built-in voices | CPU, CUDA, and supported modern AMD GPUs | | Qwen3 TTS | Built-in voices and multilingual cloning | CPU, CUDA, Vulkan, or Metal | | XTTS v2 | Mature multilingual voice cloning | CPU or CUDA | | VoxCPM2 | Large multilingual voice cloning | CUDA | | Fish S2 Pro | Broad-language voice cloning | Backend-dependent | | Voxtral 4B | Preset voices | WGPU-compatible accelerator | | Silero | Efficient regional language packs | CPU | | Chatterbox | English and multilingual cloning | CPU or CUDA | | Magpie 357M | Preset multilingual voices | CPU or CUDA | | CrispASR | Transcription, timestamps, and diarization | CPU, CUDA, Vulkan, or Apple Silicon | | RVC | Speech-to-speech voice conversion | CPU or CUDA |

Language support differs by model. Qwen3, XTTS, VoxCPM, Fish, and Chatterbox cover many common multilingual cloning workflows; Silero and Fish add especially broad regional coverage. CrispASR offers Whisper large-v3 for the broadest transcription coverage and Parakeet TDT for a smaller set of primarily European languages. Pandrator filters choices by the language and capabilities reported by each installed service.

Hardware needs vary with model size, quantization, input length, and compute backend. A GPU makes many models faster, but several useful configurations run on CPU.

Files and exports

| Category | Supported formats | |---|---| | Documents | TXT, PDF, EPUB, DOCX, MOBI, or pasted text | | Subtitles | SRT | | Audio input | AAC, AIFF, FLAC, M4A/MKA, MP3, OGG, Opus, WAV, WMA | | Video input | MP4, MKV, WebM, AVI, MOV | | Audiobook and audio output | M4B, MP3, Opus, FLAC, WAV | | Video output | MP4-oriented export with selectable audio and subtitle tracks |

URL imports use yt-dlp for supported public media sources. You are responsible for following the source service's terms and applicable law.

AI correction and translation

LLMs are optional. Pandrator can use them for subtitle correction, glossary-aware translation, document-cleaning assistance, and pronunciation or speech-text improvements. Each step creates a separate, reviewable result; AI-optimized speech text does not silently replace the text shown in exported subtitles.

Models can have their own temperature, reasoning, and token-price settings. When a provider reports an authoritative cost, Pandrator records it. Otherwise, it estimates the cost from token usage and your configured rates.

Optional Jina research can help resolve uncertain names and terminology. It uses explicit search limits and keeps a source ledger with the result.

Updating, repair, and data

Open Pandrator Manager and choose Maintenance → Check for update to update the Manager. Updates for Pandrator and local services appear as reviewable actions under Install & launch or Providers & services.

Install, update, repair, and removal actions show what will change before they run. Components are staged and checked before becoming active. User data is kept separately and is preserved by default during uninstall; deleting it requires a separate confirmation.

Closing the browser does not stop an active generation job. Reopen Pandrator from the Manager, launcher, or optional tray. Keep an independent backup of important projects before a major update.

For command-line recovery, automation, and advanced deployment details, see the Pandrator Manager guide.

Remote and headless use

Pandrator and its recovery interface listen on 127.0.0.1 by default. You can also run one Pandrator installation on a home server, a LAN or VPN host, an external server, or a GPU pod.

For an Internet-facing installation, use stable HTTPS addresses through a reverse proxy or ingress:

pandrator-manager-launcher setup \
  --workspace /srv/pandrator \
  --remote-setup-url https://recovery.example.com \
  --remote-pandrator-url https://pandrator.example.com \
  --trusted-proxy-hops 1 \
  --no-open

When the ingress runs in another container or network namespace, add --network-bind-host 0.0.0.0 and restrict the listener with the platform's network policy. Persist both the Pandrator data directory and the Manager workspace. Use a VPN or HTTPS for regular remote use; plain HTTP is available only as an explicit trusted-private-network option.

Pandrator is designed for one owner, not as a public multi-user service.

Agent access with Pandrator MCP

pandrator-mcp lets an MCP-capable agent explain Pandrator, inspect an installation, edit sessions safely, run workflow plans, and carry out bounded Manager recovery actions. The sidecar runs on the same computer as the agent and can connect to either a local installation or one fixed remote Pandrator server.

It includes secret-free configuration generators for Codex, Claude Code, OpenCode, and Antigravity. Credentials are enrolled with owner approval and stored in the operating-system credential store rather than in MCP configuration or model-visible arguments.

The Pandrator MCP guide walks through:

• local, LAN/VPN, external HTTPS, and pod targets;
• least-privilege scopes and identity pinning;
• owner-approved login and credential revocation;
• optional recovery while the Pandrator