Choosing a language pair: why one beats three

Choosing a language pair is the earliest decision in a cross-border operation and the hardest one to reverse. It determines the scripting, the product mix and the rota, and once a room is running, switching languages is effectively starting again.

Simultaneous multilingual streaming usually fails

Running Chinese to English, Chinese to Indonesian and Chinese to Vietnamese in the same session sounds like triple the reach. What normally happens is that all three audiences conclude the room was not built for them.

The constraint is not compute. It is that a script can only be designed around one audience. Some concrete collisions:

• Product selection. The same item sits in a different price band, against different competitors, with different selling points in the Philippines and in Japan. Saying “this is a discount on the usual price” means nothing to half the room, because they have no reference point.
• Payment and delivery. Cash on delivery is the default expectation across much of South-East Asia and barely worth mentioning in Japan or Korea. Covering both means everyone sits through half of something irrelevant.
• Promotions. Giveaway mechanics, thresholds and time-limited offers differ in both wording and regulation by region. Blending them tends to leave both explanations unclear.
• Pacing. Synthesised speech plays serially, so two target languages mean either double the time per sentence or one audience hearing nothing.

The workable form of multilingual operation is separate sessions sharing one technical configuration and product catalogue, with scripts written independently.

Machine translation quality is not level across pairs

Marketing lines about “100+ languages supported” tend to obscure this. In practice, usable quality falls into three tiers:

Tier	Pair	What it looks like in production
First	Chinese ↔ English	Most training data; colloquial speech and commerce terms broadly work, and errors are usually guessable from context
Second	Chinese ↔ Japanese, Chinese ↔ Korean	Strong on written register; politeness levels and casual speech slip, and product names often come out as stiff literal renderings
Third	Chinese ↔ Indonesian, Vietnamese, Thai	Noticeably weaker on short utterances and slang, with more errors in numbers, units and sizing

Be clear about what that table is: a judgement from experience, not published data. No vendor grades quality by language pair, and public machine translation benchmarks largely ignore the register this work lives in — spoken sales patter and commerce terminology. Treat it as a prior rather than a conclusion. To confirm it, use the method in the tool evaluation article: take one recording from your own broadcast, run every candidate against it, and mark the errors by hand. Half a day produces a ranking that belongs to your category rather than to anybody’s marketing.

The tiers reflect training corpus volume rather than one vendor’s weakness, so changing supplier rarely moves a pair between tiers — at best it improves things slightly within one.

The practical implication: launching your first cross-border room on a third-tier language means facing two unknowns simultaneously, translation quality and operational inexperience, with no way to tell which is causing a bad session. Getting a first-tier pair working first makes the diagnosis far cheaper.

Time zones decide whether you can keep going

A time difference is not a row on a rota. It is the question of whether the team can still do this in three months.

• South-East Asia sits within an hour of China, so peak hours overlap and normal working patterns cover it;
• Japan and Korea are an hour out, equally unproblematic;
• The Gulf is four to five hours out (UAE at UTC+4, Saudi Arabia at UTC+3), pushing evening slots past midnight;
• Europe is six to eight hours out and the United States twelve to fifteen, which means permanent night shifts. Both observe daylight saving, so the gap to a given market shifts by an hour twice a year — build the rota in a summer and a winter version rather than calculating it once.

A lot of cross-border operations collapse in week three or four, and the cause is usually not the numbers — it is a rota no human can sustain. Once the host’s energy drops, conversion follows, and the team misreads the result as “this market does not work”.

If you genuinely need Europe or North America, design for two shifts, or pre-recorded content with live interaction, from the beginning rather than asking one person to absorb two months of night work.

Native markets versus second-language markets

The Philippines, Malaysia, Singapore and India have high English penetration, so you can broadcast in English without adding a pair at all. This is usually the easiest route to a working room: first-tier translation quality, friendly time zones, and audiences already accustomed to English content.

The trade-off is worth knowing in advance. Second-language audiences tolerate complex syntax and idiom much less well. Phrasing that works for British or American viewers frequently lands flat when carried across directly. Specifically:

• Sentences need to be shorter. Processing load is higher for second-language listeners, and the translated audio already trails the original by several seconds;
• Idiom and cultural references need replacing. Expressions like “it’s a steal” or “ride or die” should become something literally transparent;
• Numbers and units need stating plainly. Repeat prices, sizes and delivery windows rather than relying on inference from context.

Conversely, native markets such as Indonesia and Vietnam offer thinner competition and more loyal audiences, but you absorb third-tier translation quality and far less published operational reference material.

A sequence you can follow directly

1. Pick one pair, preferably Chinese to English or an English-as-second-language market;
2. Run twenty consecutive sessions to stabilise scripting, product selection and templated replies;
3. Record which problems are linguistic and which are operational — only volume separates the two;
4. Once a single session breaks even, replicate to a second pair, reusing the entire technical configuration and rewriting only the script.

Going broad before going deep is almost always slower than the reverse.