Skip to main content

​
Live interpretation and mid-call language detection on voice calls

When your caller and your agent do not share a language, Orbit can interpret the call while it runs: each finalized utterance on one leg is translated into the other leg’s language and spoken back on that leg, and every interpreted segment is saved to the call record as a bilingual transcript. The softphone’s Live language detection card is where you control this mid-call: you set or override the caller and agent languages at any point, and you see what the call’s automatic language detection has observed. This guide walks through the panel line by line, in the order a bilingual support call actually goes.

​
1. What in-call interpretation is (and what it is not)

Three separate Orbit features carry “interpretation” or “language” in the name. They do not share controls:
  • Live voice interpretation (this guide). A per-call language pair on a voice call, one language per leg. While the two legs carry different languages, utterances are translated between them during the call, and each interpreted segment is appended to the call record.
  • Live interpretation in video rooms. Spoken interpreter channels in a video room, delivered on a dedicated per-language audio track. See Live interpretation in video rooms.
  • Post-call language-quality analytics. Nothing is translated; this is reporting. Containment, resolution, escalation, sentiment, and CSAT sliced by the detected conversation language, after the call. See Score AI-agent quality per language.
Everything below is about the first one, on calls handled by the browser softphone or bound to an AI agent.

​
2. Which calls can be interpreted

Live interpretation needs two things on a call: a language pair set for the two legs, and a live audio tap that observes the conversation. A call has an audio tap when it is either: If a call has neither, the panel tells you so: “Automatic detection needs a live audio tap (an AI agent or a browser-softphone room) — this call carries neither, so the language pair below is what you set.” On such a call only the pair you set manually applies; nothing detects or switches on its own. That is expected, not a fault. Two more rules of the road, both stated on the panel:
  • Same language on both legs means nothing is interpreted. When caller and agent are set to the same language, the translation round is skipped and the original audio passes through. The panel reads “Both sides are set to the same language, so nothing is being interpreted.”
  • Interpretation can be off for a call. Until you set a language, the panel reads “No call language set yet”; if interpretation is switched off for the call, it reads “Interpretation is off for this call.”

​
3. Set and override the caller and agent languages mid-call

The card carries one control per leg: an Override caller language… dropdown with a Set button, and an Override agent language… dropdown with a Set button. Both work the same way, at any point during the call. The supported languages are Arabic, German, English, Spanish, French, Italian, Dutch, Portuguese, and Turkish.
  1. On an active call, open the softphone panel and find the Live language detection card.
  2. Read the top line. This is the call’s configured pair, for example “Caller Spanish · Agent English”. This is the line the Set buttons write.
  3. Pick a language in the override dropdown for the leg you want to change, then press Set.
  4. The top line updates immediately to the new pair. The leg you did not touch keeps its current language; the very first set on a call defaults the other leg to English.
A manual Set is a straight write to the call’s language configuration. It takes precedence over whatever automatic detection did before it, and it is the pair the interpreter routes off. If you pick the same language as the other leg, the card warns you that no translation will occur and asks you to choose a different language. Interpretation only runs between two different languages.

​
4. Read the automatic detection line

Below the configured pair, a second line reports automatic detection only. When the call’s audio tap observes a leg speaking a different language, the line names the leg, the detected language, and a confidence percentage: “Caller detected speaking Arabic (87% confidence) — not switched yet.” A single detected turn changes nothing. A language must clear a confidence threshold on consecutive turns before the leg’s configured language switches on its own. When that happens:
  • the detection line ends with “switched automatically”, and
  • the card header shows an Auto-switched badge.
An auto-switch is never final. It writes the same configuration your manual Set writes, so it can always be overridden: pick a language, press Set, and your manual pair is the configured pair again. On a call with no audio tap (section 2), this line instead explains that detection needs a live tap. An empty detection line on such a call is the correct state, not a pending one.

​
5. The bilingual transcript after the call

Every interpreted segment is appended to the call record as it happens: the original text, its translation, both languages, and which leg spoke. After the call you read the conversation in both languages, line by line, on the call record alongside the regular post-call transcript; the place call records and transcripts live is covered in Call logs and call records. Transcript writes are additive and best-effort, so a storage hiccup never interrupts the live audio.

​
6. Billing

Live interpretation is metered against the same tenant translation wallet, daily caps, and rate limits as chat-message and video-caption translation. When both legs carry the same language the translation round is skipped, so a same-language call incurs no interpretation cost.

​
7. Troubleshooting

“I set a language but the panel looks unchanged.” You are reading the automatic detection line, not the configured pair. The Set buttons write the top line; the detection line only moves when the call’s audio tap reports a detection. Read the top line for what the call is actually set to. The two lines answer different questions: the top line is what you configured, the lower line is what detection observed. “Automatic detection never appears on my PSTN calls.” Automatic detection needs a live audio tap: an AI agent bound to the call, or a browser-softphone room. A PSTN call with neither has no tap, and the panel says so in plain words. The manual language pair still applies to the call; only the automatic line stays empty, by design. “The card warns me that no translation will occur.” You picked the same language as the other leg. Choose a different language for the leg you are setting, or leave both legs on one language if the call does not need interpreting.