Blog

When a Guest Speaks Another Language, the Episode Splits in Two

November 6, 2026  ·  Tablingo

A bilingual episode is not one episode with a translation problem. The transcript holds one language, the live captions hold one language, and half the conversation ends up in neither.

Your favourite show puts out an episode with a guest from São Paulo. For half the audience it is the best episode of the year. For the other half it is ninety minutes of watching two people enjoy each other's company.

So you go looking for help, and you find something stranger than you expected: the episode already has a transcript. It is just not reading right.

That is the tell. A bilingual episode is not a normal episode with a language problem attached to it. It is two episodes sharing one audio file, and every layer the podcast industry built on top of that file — the transcript, the live captions, the chapters, the clips, the search box — was built to hold one language. Each layer resolves the ambiguity in its own way, and none of them tells you which half it kept.

The platform says it out loud

YouTube publishes the conditions under which its automatic live captions can miss speech. Poor audio, unclear speech, overlapping speakers, and — this is the one that matters here — multiple languages spoken at once. The platform's own troubleshooting guidance names the exact format this article is about.

The same guidance settles the rest of the terms, and they are worth reading together: those captions are available in English only, they require normal latency, they are configured per stream rather than across a channel, and once the broadcast ends they do not stay on the video. A separate track is generated for the replay, and it can differ from what viewers saw live.

Put those two paragraphs side by side and the situation is plain. The format most in need of captions — two people speaking different languages — is the one the platform's captions are documented not to handle. And the half that does get handled is whichever half is in English.

A transcript has a language, and it has one

Apple's documentation for podcast creators lists the languages it generates transcripts in and the languages it accepts from you, and the two lists are the same: English, Danish, Dutch, Finnish, French, German, Italian, Norwegian, Portuguese, Spanish, Swedish. Eleven. Transcripts have been generated automatically since iOS 17.4 and reach listeners in more than 170 countries and regions.

It is tempting to read that as "my language might not be supported." That is true for Japanese, Korean or Mandarin shows, and it is not the interesting part. The interesting part is that a transcript is generated in a language — once, for the episode, from the audio. An episode that is two of those eleven languages is not one of the eleven. There is no row in that table for it.

What happens next is worse than nothing. The recogniser does not return a blank for the guest's half. It returns its best guess, rendered in the language it was primed for. Fluent text, properly spelled, saying something the guest did not say. If you can hear the audio you can spot it. If you were reading precisely because you cannot, you have no way in.

Two audiences, one text layer

This layerSpeaks the guest's languageDoes not
Live audioThe conversationCadence and nothing else
Live captionsEnglish onlyEnglish only
TranscriptReadableThe recogniser's guess

Read the last column first. The only listener who gets the whole episode is the one who never needed the tool.

Note also that "English only" appears in both columns. The rule is not that each audience gets its own language. The rule is that the layer holds one, and English is the one it holds. A Spanish-speaking listener is exactly as far from that transcript as you are.

Why every layer decides once

None of this is laziness. A speech recogniser is not the language-neutral thing people assume it is. Give it a language and it sharpens on that language — vocabulary, phonetics, punctuation habits, the shape of a sentence — and it dulls everywhere else. So every practical system has to commit, and the commitment is usually made early and then held for as long as the session lasts.

Our own standard mode is a fair example, and it is ours to describe. Leave the source language on auto and the first chunks are detected. When two consecutive results agree, that language is locked for the rest of the session. From that moment every chunk is sent with two things attached: the locked language, and the text already produced, as a prompt. The second one matters more than it looks. The transcript so far keeps pulling the next minute toward the language it is already written in.

So if the guest switches language at minute twenty, the pipeline is by then heavily invested in the host's language. It does not stop. It keeps returning fluent English that no longer resembles what was said — which is why this failure is easy to miss and expensive when it happens.

Following the switch instead of choosing

The real-time path does not make that commitment. It takes hints rather than a verdict, and if you leave the source on auto it takes no hint at all: the model decides token by token across the whole session. A switch in the middle of an episode gets followed rather than flattened.

The visible result is small, and once you have seen it, hard to give up. The guest's line comes back in the guest's language, and the translation into your language sits underneath it — because the translation runs off what was actually recognised, not off a language somebody chose before the show started. On the host's lines the top line is English. On the guest's lines it is not. The seam stays visible, which is the entire point.

The boundaries deserve the same plainness. Speaker labels are opt-in and exist only on that path, so a two-voice conversation arrives as one subtitle stream rather than two labelled columns. A speaker who changes language inside a single sentence is harder than one who changes between sentences. And all of it happens inside a browser tab, so a podcast playing in a desktop app is outside the boundary — a real cost for this subject, because a good deal of podcast listening happens in exactly those apps.

Four routes, and when each one decides

RouteAvailableHow language is decided
Platform transcriptAfter publicationOnce, by the episode
Platform live captionsOnly while liveEnglish, always
Post-hoc transcription serviceOnce a file existsOnce, by you
Reading as it playsDuring playbackContinuously, per line

Three of the four commit before the episode begins. Only the last one keeps deciding, and that is the whole difference for an episode that changes language halfway through.

The routes are not competing on quality. A platform transcript can beat anything you generate yourself, if the episode is in a language it covers. The question is narrower: does the route still work on the half of the episode that is not in the show's main language.

Two checks, five minutes

  1. Play the guest's first minute with captions on. If the text is fluent and does not match the words you can hear, the recogniser was primed and is guessing. If it is empty, it declined — which is at least honest.
  2. Open the transcript and find the seam where the guest starts. If the text turns confident and wrong at exactly that point, the whole text layer is one language deep, and re-reading will not recover the other half.

Neither check needs an install or a vendor's word, including ours.

What this leaves

An interview is, by design, somebody from somewhere else. That makes the bilingual episode the ordinary case rather than the exotic one, and it makes it the hardest test in this category — because it removes the assumption every tool would like to keep, that the interesting part will be in the same language as the rest of the show.

A tool that needs one language for a whole session can be excellent and still be the wrong tool for this episode. That is not a ranking. It is a shape, and shapes have edges.

What you want is smaller than a feature list: you should never have to know in advance which language the interesting part will be in.

If you want to try ours, it's called Tablingo — a browser extension that reads the tab you are already playing and puts the original line above the translation. The general podcast case is here, the live-stream case here, and how the two capture strategies differ here. Why the recogniser has to decide which language you are speaking before it writes a word is here.