Live Translate
1. What it is
Live Translate is a real-time speech-to-speech interpreter for frontline conversations. Two people who do not share a language each speak in their own, and the app transcribes, translates, and speaks back in the other person’s language — no bilingual colleague needed.
- Enablement: Live Translate is pay-as-you-go. It requires no license; usage meters through the AI/Voice credit pool. An admin enables it from the Apps Marketplace.
- What it is not: Live Translate is the real-time spoken interpreter. The company glossary it shares also governs written translation of voice notes, recognition posts, alerts, policies, and Sites — but those surfaces are part of their own apps, not this one.
2. Standing it up
- Go to Admin → Apps Marketplace, find Live Translate, and enable it.
- Open Live Translate in Admin → Apps Marketplace and click Publish to users — or Publish to specific groups. Enabling alone doesn’t give employees and managers access.
- (Optional) Go to Admin → Organization → Integrations and connect a dedicated translation provider — AWS Translate or DeepL. Without one, translation falls back to the platform-managed engine automatically; no extra configuration is required.
- (Optional) Go to Admin → Translation and adopt an industry starter pack or add your own glossary terms so company-specific vocabulary translates correctly from the start.
- Confirm it works: open Live Translate from the app launcher, select two languages, hold the talk button, and speak. You should hear the translation spoken back.
Admins can use the app as soon as it is enabled; employees and managers can once it is published to them (or their group is given access). The glossary and the dedicated provider are governance layers — they improve accuracy, but nothing blocks use without them.
3. How it fits together
Two interpreter modes
Live Translate offers two ways to use it, One device and Two devices, listed in the app’s side menu. On a phone they also appear as buttons at the top of the page.
One device — two people share a single phone or tablet. Each side of the screen has its own language selector and talk button. Person A holds their button and speaks; the translation plays aloud for Person B, and vice versa. Example: a warehouse supervisor (English) and a new hire (Spanish) stand together reviewing a safety checklist — one phone between them, each holding their button to talk.
Two devices — each person uses their own phone. One person starts a session and receives a six-character code; the other enters it on their device to join. Once paired, each person speaks on their own screen. The translation is broadcast to the other device and played automatically. Two devices also supports a hands-free mode where the app listens continuously and sends each utterance for translation when the speaker pauses, with no button press required.
Live Translate isn’t in the MangoApps mobile app — on a phone, open it in the phone’s web browser: sign in at your company’s MangoApps address followed by /m/login, then open Live Translate from More (or go to /apps/live-translate).
A paired session expires after 2 hours. Only the two participants can send or receive turns — a third person with the code is rejected.
Using a screen reader, voice control or switch access? Press Hold to speak once to start and again to stop.
The turn pipeline
Each spoken turn runs three steps in sequence: the speech is transcribed (speech-to-text), the transcript is translated through the resolved provider, and the translation is spoken back (text-to-speech). Turns are ephemeral: Live Translate keeps no conversation record, there is no transcript history to open, and nothing you said is readable in the product afterwards — except a phrase someone flags as mistranslated (see The quality loop below). The transcript and translation appear on screen only while the conversation is open.
One qualification, stated plainly because “never stored” would be too strong. In two devices mode each turn travels to the other phone over MangoApps’ shared real-time message service, labelled with your organization’s ID, and is deleted within roughly 90 minutes by a cleanup job that runs every 30 minutes. Until then the turn’s text and its synthesized audio are held there. No screen in the product reads them; reaching them requires direct database access. One device mode does not use the message service at all. This is delivery, not retention: nothing accumulates, and there is still no conversation history anywhere in the product.
The company glossary
The glossary is a set of terminology rules managed at Admin → Translation. Besides Live Translate, it applies to written translation elsewhere on the platform, including voice notes, recognition posts, alerts, policies, and Sites. Chat message translation, Service Desk translation, broadcast voice calls, and SOP Hub’s AI translate don’t apply it yet.
Two types of rule are available:
- Mapping — a term in one language must translate to a specific term in another. For example, “forklift” (English → Spanish) must always become “montacargas”, not a generic translation a provider might choose.
- Do not translate — a term must pass through untranslated regardless of target language. Brand names, product codes, and acronyms like “OSHA” are common examples.
A worked example: a retail chain adds three glossary rules — “stockroom” → “almacén” (mapping, English → Spanish), “backstock” → “existencias de reserva” (mapping, English → Spanish), and “SKU” (do not translate). When a supervisor says “check the backstock SKU in the stockroom”, the Spanish translation uses “existencias de reserva” and “almacén” and keeps “SKU” as-is — for example “revisa el SKU de existencias de reserva en el almacén” — instead of whatever words the engine would have picked. The same three rules also apply when an English voice note is translated into Spanish.
Terms match whole words — add plurals as their own terms. A mapping applies only when a translation runs in exactly its language pair.
The glossary enforces rules at the platform layer, not inside any one provider. The mechanism is protect-then-restore: matched terms are replaced with placeholder tokens before the text reaches the translation engine, and the correct terms are substituted back after translation returns. This means rules work identically whether the tenant uses AWS Translate, DeepL, or the platform’s LLM engine, and whether the surface is the spoken interpreter or a written translation.
The quality loop
Any user can flag a translation as wrong by tapping the Wrong? button on a translated turn (or on a translated voicenote). The flag lands in a pending queue visible at Admin → Translation. An admin can then:
- Add term — type just the word or short phrase that came out wrong into Term (narrow to the word) (the flagged sentence is on the card as Original:), enter its correct translation in Should translate to, and click Add term. Keep it short: a term matches only that exact wording, so a whole sentence rarely matches a later turn.
- If the card says a term can’t fix the flag — the speaker’s language wasn’t recognised, or both sides were set to the same language — there’s nothing to add; Dismiss it once you’ve read it.
- Dismiss it — close the flag if the translation was acceptable. Dismissing doesn’t delete it: its phrase still counts in the most-flagged phrases on the Analytics page.
Admins receive an in-app notification when flags arrive (throttled to one per 6 hours) and an email digest (throttled to one per 24 hours).
Provider resolution
Translation runs through the first configured engine in this order:
- A dedicated provider the admin connected at Admin → Organization → Integrations — AWS Translate or DeepL (bring-your-own-key).
- The platform-managed engine — AWS Translate with platform credentials, available to every tenant with no setup.
- The platform’s language model — used only where neither of the above is set up.
The order is decided by what is set up, not by outages: if the chosen engine fails during a spoken turn, that turn fails and the speaker tries again. It never switches to another engine mid-conversation.
Speech-to-text and text-to-speech always run on the platform’s OpenAI key. If that key isn’t set up, Live Translate shows a notice on load and disables the talk buttons, rather than letting users discover the failure mid-conversation. If text-to-speech fails on a single turn, the translation still appears on screen.
Supported languages
Live Translate supports 15 languages: English, Spanish, French, German, Portuguese, Italian, Chinese, Japanese, Korean, Arabic, Hindi, Russian, Vietnamese, Tagalog, and Polish.
In one-device mode, either side can also be set to Detect language, which identifies the speaker’s language automatically from their speech. Both sides cannot be set to Detect at the same time — the Detect side must speak first so the app knows what language to translate into.
Your side — Person A on one device, Your language on two devices — starts on your profile language (English if your profile language isn’t one of the 15). On one device, Person B starts on Spanish, or English if your profile language is Spanish. Each device remembers the last languages you picked.
4. Running it
Managing the glossary
- Go to Admin → Translation.
- Add terms individually using the form at the top, or adopt an industry starter pack. Three packs ship with the platform:
- Warehouse & Manufacturing Safety (English → Spanish) — 13 terms including lockout-tagout, PPE, and muster point
- Retail & Store Operations (English → Spanish) — 9 terms including stockroom, planogram, and shrinkage
- Healthcare & Facilities (English → Spanish) — 7 terms including shift handoff, charting, and isolation room
- To bulk-add terms, import a CSV file with the header row shown below. Re-importing a previously exported file skips duplicates automatically.
- To export the full glossary, click Export CSV — useful for backup or for sharing terminology across tenants.
The file uses the same columns Export CSV writes, under this header row:
source_language,target_language,source_text,target_text,term_type,case_sensitive,active
term_type is mapping or do_not_translate. The last two columns are optional — leave them out or blank and a term imports as case-insensitive and active; otherwise put true or false. Save the file as CSV UTF-8 with commas between columns.
Each term can be marked active or inactive, and optionally set to case-sensitive matching. Language codes use ISO 639-1 format. For Live Translate, use two-letter codes such as “en”, “es”, and “pt” — its turns never use a regional code, so a term written for “pt-br” doesn’t apply there, though regional codes like “pt-br” still apply on other translation surfaces.
Reviewing flagged translations
- Go to Admin → Translation. The Flagged translations section shows pending flags, oldest first.
- For each flag, decide whether the translation was genuinely wrong.
- To fix it, type the offending word into Term (narrow to the word) — it starts empty — enter its correct translation in Should translate to, and click Add term — this creates a glossary mapping and closes the flag in one step.
- To close a flag without action, click Dismiss.
The flagged-phrases list on the Analytics page shows which phrases are flagged most often — dismissed and resolved flags included — and whether the glossary already covers them.
Viewing translation analytics
- Go to Admin → Translation, then click Analytics.
- The dashboard shows counts of machine translations (calls to AWS Translate or DeepL — translations written by the platform’s language model aren’t counted), voice transcriptions and spoken playbacks (both cover every feature, not only Live Translate), and active terms; a daily trend chart; the active engine with average latency, cached translations and a count per provider; the terminology quality loop (pending, resolved, dismissed flags); top target languages; and the most-flagged phrases. The window is selectable at 7, 30, or 90 days; AI usage logs are kept for 30 days, so a 90-day window shows no activity before that.
Analytics show counts and languages, plus the most-flagged phrases table, which shows the text of phrases people flagged (all time, dismissed and resolved flags included). That table cross-references each phrase against the active glossary, so you can see at a glance which recurring complaints already have a term covering them and which still need one.
5. Settings
Live Translate has no per-app configuration keys. Its behavior is governed by:
| Setting | Where | Default | What it changes |
|---|---|---|---|
| App enablement | Admin → Apps Marketplace | Off | Adds Live Translate to your app list. Once enabled, it is admin-only until you publish it or give groups access (next row). Live Translate is pay-as-you-go, so the app cannot be switched off from here |
| Published to users | Admin → Apps Marketplace → Live Translate → Access & Permissions | Off | Publish to everyone lets every employee and manager open Live Translate; giving specific groups access limits it to those groups instead. With neither, only admins and people you add as the app’s admins, contributors, or collaborators can open it |
| Translation provider | Admin → Organization → Integrations, category Machine Translation (AWS Translate / DeepL) | Platform managed | Which engine handles the translation step |
| Glossary terms | Admin → Translation | None | Company terminology applied to Live Translate turns and the other surfaces that use the glossary (see The company glossary) |
Glossary terms default to Active and case-insensitive matching when created. Flagged translations stay Pending until an admin adds a term from them or dismisses them.
Sample data
To explore the glossary and analytics without creating real terms, go to Admin → Translation and click Load Sample Data. This creates demonstration glossary terms and flagged suggestions you can experiment with, then remove when you are ready.
6. More help
- Live Translate FAQ — specific setup and operating questions, limits, and troubleshooting
- Ask AI — the assistant answers questions about Live Translate from these articles.