Make audiobooks, subtitles, and voiceovers in one workspace.
Turn a book into something you can listen to. Transcribe a recording, translate its subtitles, or give it a new voice. Pandrator brings the steps together in a browser interface, with room to review and refine the result as you go.
Run speech and transcription models on your own computer, connect a cloud provider, or let your AI assistant help through MCP. Start with one workflow and add more when you need them.
Windows 10/11 and Linux desktop · x86-64 · Pandrator 0.9.4 / Manager 0.9.24
All downloads & release notes · Installation help · User guides
- Download and run Pandrator Manager. On Linux, make the AppImage
executable first. Choose a parent folder; the Manager creates a
Pandratorworkspace inside it and opens the setup interface in your browser. - Install Pandrator and the engine you need. Try Kokoro for ready-made narration voices, or CrispASR for transcription. You can add voice cloning and other engines later. Docker, WSL, and a separate Python installation are not required.
- Make something small. Choose Generate an audiobook and paste a short passage, or open Quick Transcribe to upload a clip or record your microphone. Listen, review, and save the result.
Local models download separately on first setup; speed and memory requirements vary by engine and hardware. The Manager shows the available compute options. The Windows launcher is currently unsigned, so Windows may show an unknown publisher warning; the installation guide explains this and the release checksums.
| Start with… | Make… | Guide |
|---|---|---|
| A book, document, or pasted text | Narrated audio or an M4B audiobook with chapters and cover art | First audiobook |
| An audio/video file or existing subtitles | Reviewed, corrected, or translated subtitles | First subtitles |
| A video or subtitle file | Synchronized speech and a dubbed video | First voiceover |
| A recording and optional captions | Transcript-guided cuts, reviewed boundaries, and an edited video | Recording editing with MCP |
| A short clip or your microphone | A TXT, SRT, or JSON transcript without creating a session | Quick Transcribe |
If you use Codex, Claude Code, OpenCode, or another MCP-capable host, its model can do Pandrator's language work directly. You do not need to configure a separate LLM provider or API key in Pandrator for this route.
This is the in-harness passive MCP workflow: Pandrator prepares and tracks the work; the model in your existing conversation processes it and submits results through MCP. “Passive” describes Pandrator's role: it does not make the LLM calls itself.
flowchart TD
P["Pandrator prepares a batch"] --> A["Your agent corrects, translates, or optimizes"]
A --> V["Pandrator validates and saves"]
V -->|More batches| P
V -->|Complete| R["Review and export, or generate speech"]
Use it to:
- Clean PDF/EPUB text before audiobook preparation, with reviewable changes to document structure and extraction artifacts.
- Correct and translate subtitles, keeping cue identities and shared terminology across batches.
- Optimize text for speech before narration, while retaining the source wording as a separate artifact.
- Coordinate a whole workflow: import a file, process the text, select a voice, generate speech, and return verified deliverables.
After connecting your MCP host, try a request such as:
Use Pandrator to correct this recording's English subtitles and translate them into Polish. Use your own model through passive dispatch, keep a shared glossary, and export both SRT files. Preserve uncertain wording for review.
Pandrator tracks batches, validates submissions, and keeps completed work so an interrupted agent can resume. The agent must continue the tool loop; starting a passive run alone does not process it. Speech recognition and speech generation still use the engines you select in Pandrator.
Your host's normal usage limits and costs apply. If its model runs in the cloud, the text it processes is sent there, even when Pandrator runs locally.
See the passive workflow guide, end-to-end agent guide, and the reusable workflow skill. The skill adds workflow guidance; the MCP connection supplies the tools. See skill installation to add it to your host.
- Import TXT, PDF, EPUB, DOCX, MOBI, or pasted text. MOBI conversion needs optional Calibre.
- Prepare documents with extraction, OCR, chapter detection, text cleanup, and optional AI assistance.
- Adjust pronunciation and speech text, generate in segments, compare takes, and regenerate selected passages.
- Record or upload voice references and retain their transcripts.
- Export WAV, MP3, Opus, FLAC, or M4B, with audiobook chapters, metadata, and cover art where the format supports them.
- Transcribe audio/video with timestamps and optional speaker diarization.
- Import SRT; WebVTT, ASS, and SSA uploads are also recognized, with SRT used as the working subtitle format.
- Correct, translate, split, merge, and retime cues in a reviewable editor. Use manual edits, configured language providers, or passive MCP processing.
- Generate synchronized speech and export media with selectable audio and subtitle tracks, including burned-in subtitles.
- Plan recording cuts from a transcript, inspect and refine their boundaries, and render a reviewed edit while preserving the original media.
- Upload audio/video or record and preview your microphone, then copy or download TXT, SRT, or JSON without creating a permanent session.
- Transcribe through HTTP or MCP: an approved-root file path or a small base64-encoded clip can be supplied through MCP. Larger files are streamed from the host rather than placed in the model's context.
- Quick transcripts are temporary and expire one hour after completion; save the result you want to keep.
- Use MCP for session workflows, live voice/model selection, reviewable plans, resumable work, and downloads verified by size and checksum.
audio.cpp is a local speech provider with several models to choose from, and the default for new workspaces. Select the model packages you want to install in the Manager, then choose an installed model in Generate audio settings. You can switch models or add more later. Available families include:
| audio.cpp model family | Voice options |
|---|---|
| Qwen3 TTS 1.7B | Base for reference-voice cloning; CustomVoice for built-in speakers; VoiceDesign to describe a new voice in any of ten supported languages. |
| Fish Audio S2 Pro | Expressive reference-voice cloning. |
| VoxCPM2 | Multilingual reference-voice cloning. |
| Chatterbox | Reference-voice cloning. |
| MagpieTTS | Preset multilingual voices. |
| OmniVoice | Multilingual reference-voice cloning. |
| PocketTTS English | English speech generation. |
| FireRedTTS3 Base | Experimental multilingual reference-voice cloning. |
| BreezeTTS 2 | Instruction-based voice design and optional reference cloning in English and Mandarin Chinese. |
The Manager supplies selectable Q8_0 model packages. CPU, Vulkan, and CUDA builds are available; the Linux CUDA build is currently best-effort and has not been verified on NVIDIA hardware. Available voices and languages depend on the selected model.
These dedicated providers also remain available:
| Provider | Voice options and capabilities |
|---|---|
| Kokoro-82M v1.0 | Lightweight, CPU-friendly preset voices; a useful first narration engine. |
| Silero | CPU-friendly language-specific voice packs; licence terms differ by pack. |
| XTTS v2 | Multilingual and cross-language voice cloning; import fine-tuned model bundles or use the optional XTTS training component. |
| Voxtral | Preset voices; requires a supported GPU/WGPU backend. |
Existing workspaces keep their saved defaults and provider selections. Standalone Qwen3 TTS, Fish S2 Pro, VoxCPM2, Chatterbox, and Magpie providers remain supported for compatibility. Installed components stay manageable; uninstalled ones appear under Compatibility backends. In a saved session, Switch to audio.cpp proposes a matching model family for review. Voices and engine settings are not assumed interchangeable, and existing takes remain intact. See switching providers.
Optional RVC converts the voice of generated audio, keeping the original and converted takes available. It is a separate post-processing step. XTTS training benefits strongly from an NVIDIA GPU.
| Runtime / provider | Models and capabilities |
|---|---|
| CrispASR (local) | Whisper large-v3 for multilingual transcription, Parakeet TDT 0.6B v3 with native word timing, and MOSS Transcribe-Diarize 0.9B with speaker diarization. |
| Canary CTC (local) | Forced alignment for existing captions and MOSS word timing; aligns text to audio. |
| Azure Speech (cloud) | MAI-Transcribe-2 and MAI-Transcribe-1.5, with word timing; these profiles do not provide diarization. |
- OpenAI:
gpt-4o-mini-tts,tts-1, andtts-1-hd. - Google Gemini / Vertex AI: Gemini 3.1 Flash TTS Preview, 2.5 Flash TTS, and 2.5 Pro TTS.
- ElevenLabs: native API integration with live model and voice discovery.
- Azure Speech: MAI Voice 2 and MAI Voice 2 Flash preset voices.
- External servers: OpenAI-compatible and generic JSON endpoints, with profiles for services such as Piper, StyleTTS2, Matcha-TTS, and Open Unified TTS, plus Azure OpenAI TTS deployments. These are connections to separately operated services; they are not all Manager-installable engines.
Cloud access depends on your provider account and available models.
Use OpenAI, Anthropic, Google Gemini, OpenRouter, Groq, Mistral AI, Azure OpenAI, Google Vertex AI, or Amazon Bedrock, or connect a local LM Studio, Ollama, or OpenAI-compatible server. Model selection follows the provider's catalogue or your configured model ID.
Correction, glossary-aware translation, document cleanup, and speech-text optimization can use a configured LLM. DeepL is also available for translation. The passive MCP route above lets your existing host model do language work instead.
The providers and voices guide covers setup, compute choices, voice references, and model imports.
Local speech and transcription engines can keep processing on your own machine. Cloud speech, language, and translation services are optional; their selected inputs are sent to those providers. Model language coverage, hardware requirements, and licences differ. Check the engine's details before a long run.
Pandrator is designed for one owner and listens locally by default. You can also run it on a home server or GPU host; follow the remote and headless guide for setup and the privacy guide for data flows.
| Topic | Documentation |
|---|---|
| Setup and step-by-step workflows | Documentation index |
| Engines, voice references, and model imports | Providers and voices |
| PDF/EPUB extraction and cleanup | Document ingestion |
| Correction, translation, and glossaries | Language workflows |
| Pronunciation and narration preparation | Pronunciation · Speech optimization |
| Supported output formats | Formats and exports |
| Installation, updates, and repair | Manager guide |
| MCP connection and workflow skill setup | MCP guide |
For source development, install Pixi and use the committed lockfile:
git clone https://github.com/lukaszliniewicz/Pandrator.git
cd Pandrator
pixi install --locked
pixi install --environment web-build --locked
pixi run --environment web-build web-build
pixi run serve-webRun pixi run run-worker in a second terminal. See
development from source and
contributing for the development workflow
and focused checks.
Stuck on setup? Start with troubleshooting. Found a bug or have an idea? Open an issue with your operating system, engine/provider, and steps to reproduce it. For installation or service failures, the Manager's Download diagnostics bundle can help; inspect it before sharing.
Pandrator's source code is available under the MIT License. Third-party models, dependencies, and providers have their own licences and usage terms. Check those separately, and use source media and voice references you have the rights to use.
