Feedback

Stagetab grows with its users: the bugs you report and the tools you ask for help decide what comes next.

The tool, version and browser are sent along so the bug can be reproduced. Your email only if you enter it.

Most requested tools

  1. Loading …

Click ▲ to vote. New requests appear after a short review.

← All tools

Live translation

Speech from the desk becomes text in real time and is translated into other languages – as subtitles in the tool, on a screen or projector and as a file. Everything runs on this computer: audio and text never leave it.

Machine translation – may contain errors. Audio and text stay on this computer.

Guest phones

Guests read the subtitles on their own phone: scan the QR code, choose a language. The hardware bridge provides a separate, read-only page on the network for this – phones and this computer must be on the same network (guest Wi-Fi with “client isolation” prevents this). No internet needed, nothing leaves the network.

With sound, guests hear the computer voice of their language – like an interpreting system, with their own headphones. About 64 kbit/s per phone while speech is playing. Without a local speaker: choose “guest phones only” as output under “Speech output”.

Speech output

Reads the translation aloud with a computer voice – each language on its own audio output, e.g. a channel of the interpreting system or a headphone transmitter. Voices: Piper (rhasspy/piper-voices), pronunciation: eSpeak NG. This module is licensed under GPL-3.0 and kept separately in the AI folder (tts/), including source code. The voices are loaded once like the other models.

Read aloudLanguageOutputVolumeVoice

Labelling under the EU AI Act (Art. 50): the voice is AI-generated. Guest phones show “AI voice” and every audio frame carries a machine-readable marker; for loudspeakers and interpretation systems there is the announcement. Please have this confirmed legally before use.

Settings
Translate to

One line per entry. Just a name (e.g. ES-Sound): hint for speech recognition, never translated. “wrong = right”: corrects the recognised text. “> word = replacement”: replaces in the translated text.

Subtitle output

Models

The AI models are downloaded once at the push of a button and stay stored in this browser; after that everything runs without internet. Source: the own mirror on the Stagetab server, otherwise Hugging Face. Every file is checked against a fixed checksum (SHA-256) and discarded otherwise.

ModelQuantityLicenceRevision

How it works

Signal path: An aux or matrix output from the desk (speech microphones only, no music or playback) goes into an audio interface on the computer. Choose input and channel here and press “Start”. A laptop microphone is fine for trying it out, but not reliable in the venue.

How it works: Voice activity detection (Silero VAD) cuts out the spoken passages. Whisper transcribes about every second (grey = preliminary). Text only counts as confirmed once two passes agree; after a pause or at the latest after 10 s the passage is closed. Only confirmed text is translated, at sentence and clause boundaries (Opus-MT, one model per language pair).

Languages: Spoken language is German or English. Translation targets are German, English, French, Spanish, Italian, Dutch, Ukrainian and Russian, with “quality” also Polish, Turkish and Japanese. Where there is no direct model from German, the translation goes via English (marked) – this costs some accuracy and time. With “Translation: Qwen3.5” a small language model (2 billion parameters, 1.4 GB, GPU only) translates all languages directly and usually more fluently. It needs more time per sentence and, like any language model, can make mistakes; Polish, Turkish and Ukrainian are still somewhat rough. All languages except Turkish and Japanese can be read aloud – there is no freely usable voice for those yet.

Computer: For “Whisper large-v3-turbo” a graphics card with WebGPU (NVIDIA or AMD, at least 8 GB video memory) in Chrome, Edge or the desktop app. Without WebGPU the processor does the work; then choose “Whisper small” or “base” – they are less accurate and slower.

Delay (honestly): with a graphics card usually 1.5–3 s from speech to translation, without one 3–8 s. The display at the top shows the measured value. It does not replace interpreters, but for subtitles and accessibility it is often enough.

Privacy: Audio, transcript and translation stay on this computer. The only network traffic is the one-time model download. The project file contains only settings, languages, glossary and output layout – never audio or text.

About Stagetab

All calculations run locally in your browser. Nothing you enter is transmitted; run sheets and soundboard labels are stored on your device only.

The calculators provide guide values for planning and setup. Manufacturer data, on-site measurements and safety regulations always take precedence.

Need technical production for your event?

ES-Sound plans and runs sound, lighting and video for corporate events, concerts and celebrations.

Request a quote

Licences & trademarks

Terms of use

  1. Stagetab may be used free of charge for planning and running your own events, including commercially.
  2. Redistribution, sale, modification, reverse engineering and copying parts of the software into other products are not permitted. Stagetab is provided exclusively via es-sound.de.
  3. All calculations (e.g. loads, power, levels) are guide values and do not replace checks by competent personnel. The user or the responsible specialist on site is responsible for set-up, safety and sign-off.
  4. Stagetab is provided without warranty. ES-Sound is only liable for intent and gross negligence, as far as permitted by law.
  5. Project data stays in your browser or on your computer. Stagetab does not send any project data to ES-Sound.

Software, data and fonts

The ES-Sound Stagetab uses the following software, data and fonts. The full licence texts are included with the files.

ComponentRights holderLicence
three.js (3D rendering)three.js authorsMIT
systray (hardware bridge tray icon)Fyne.io project, GetlanternApache 2.0
godbus (hardware bridge on Linux)Georg Reinke and contributorsBSD 2-Clause
Go x/sys (hardware bridge on Windows)The Go AuthorsBSD 3-Clause
MapLibre GL JS (map in the site plan)MapLibre contributorsBSD 3-Clause
Map data in the site plan (only when a map is selected)OpenFreeMap, OpenMapTiles, OpenStreetMap contributorsODbL
PDF.js (PDF as floor plan)Mozilla FoundationApache 2.0
ONNX Runtime WebMicrosoft CorporationMIT
AI model “Beat This!” (beat tracking)Institute of Computational Perception, JKU Linz; Johannes GorsetMIT
Essentia models for mood detection (AI module only)Music Technology Group, Universitat Pompeu Fabra, BarcelonaCC BY-NC-SA 4.0
fft.jsFedor IndutnyMIT
Transformers.js (live translation, in the AI module)Hugging Face, Inc.Apache 2.0
Silero VAD 6.2.3 (voice activity detection, in the AI module)Silero TeamMIT
AI model Whisper large-v3-turbo (only after download in live translation)OpenAI; ONNX version onnx-communityMIT
AI models Whisper small and base (only after download)OpenAI; ONNX version onnx-communityApache 2.0
AI models Opus-MT English → German (only after download)Language Technology Research Group, University of Helsinki (Helsinki-NLP); ONNX version XenovaCC BY 4.0
AI models Opus-MT German → English/French/Spanish, English → French/Spanish/Italian/Dutch/Russian/Ukrainian (only after download)Language Technology Research Group, University of Helsinki (Helsinki-NLP); ONNX version XenovaApache 2.0
AI language model Qwen3.5-2B (“quality” translation, only after download)Qwen team (Alibaba Cloud); ONNX version onnx-communityApache 2.0
Speech output of live translation: eSpeak NG (separate module in the AI folder, with source code)eSpeak NG contributorsGPL-3.0-or-later
AI voices Piper (speech output, only after download): Thorsten (de), Davefx (es), Pim (nl), ukrainian_tts (uk), Darkman (pl)Thorsten Müller; Open Home Foundation (rhasspy/piper-voices)CC0
AI voice Piper LJSpeech (en, only after download)Keith Ito (LJ Speech), Bryce Beattiepublic domain
AI voices Piper Siwis (fr), Paola (it), Dmitri (ru) (only after download)SIWIS (University of Edinburgh); Open Home Foundation; base model hifi_tts_92 from the Hi-Fi Multi-Speaker English TTS Dataset (NVIDIA Corporation)CC BY 4.0 (Siwis, base model), CC0 (Paola, Dmitri)
Open Fixture Library (fixture data)Florian & Felix Edelmann and contributorsMIT
GDTF files from GDTF Share (Powered by GDTF)Manufacturers or users who uploaded the file; GDTF Share by MA Lighting, Robe and VectorworksTerms of use of gdtf-share.com; GDTF is DIN SPEC 15800
Fonts Barlow Condensed, IBM Plex Sans, JetBrains MonoJeremy Tribby; IBM Corp.; JetBrains s.r.o.SIL Open Font License 1.1
Windows app: Electron and ChromiumOpenJS Foundation, The Chromium AuthorsMIT and others, see program folder

GDTF and MVR are open formats according to DIN SPEC 15800. The lighting plot symbols follow the USITT recommendations (RP-2) and are our own drawings.

Trademarks

Product and company names mentioned, e.g. of fixture, truss and console manufacturers, are trademarks of their respective owners. They are used only to identify devices, formats and compatibility. The Stagetab is not affiliated with or endorsed by these manufacturers. Manufacturers' fixture files remain their property.

Live translation turns speech from the mixing desk into text in real time and translates it, starting with German and English. You see the subtitles in the tool, on a second screen or projector and as a file. Audio and text stay on your computer, because the AI models run right in the browser or the desktop app.

Live translation with original text and English subtitles in the browser

Getting the audio into live translation

Send an aux or matrix mix with the speech microphones to an audio interface on the computer. Keep music and playback off this mix. In the tool you choose input and channel and press Start. Voice activity detection cuts out the spoken passages, Whisper transcribes, and only confirmed text is translated.

Subtitles on projector, screen and stream

The subtitle output opens full screen on the screen you choose. You set language, number of lines, font size and position per output. With chroma green as background you key the text over the picture in your vision mixer. After the event you save transcript and translation as TXT or SRT.

Computer and delay

Live translation works best with a graphics card with WebGPU and at least 8 GB of video memory. The delay is then usually 1.5 to 3 seconds. Without a graphics card the processor runs smaller models, slower and less accurate. Machine translation can contain errors and does not replace interpreters, but for subtitles and accessibility it is often enough.

Together with other tools

The subtitles go well with the video mixer, which mixes the picture for projector and LED wall. Check the level of the speech mix beforehand with the RTA analyser, and the run sheet shows when each speech starts. Speech recognition is based on Whisper by OpenAI, translation on Opus-MT by the University of Helsinki. You download both models once at the push of a button, and every file is checked against a fixed checksum. After that, live translation works without any internet at all, for example in a hall without Wi-Fi.

Nach oben scrollen