---
title: "Four approaches to live translation, compared"
description: "Platform built-ins, desktop tools, self-hosted pipelines and live interpreters compared under stated conditions, with our own disclosure."
canonical: https://obstrans.net/en/blog/live-translation-approaches-compared
language: en
topic: tool-reviews
published: 2026-08-06
keywords: live translation approaches, platform built-in translation, self-hosted translation pipeline, live interpreter, vendor selection
---

# Four approaches to live translation, compared

> There is no best live translation setup, only one that fits a given language pair, budget and latency tolerance. This site is maintained by the OBSTrans team, which builds a desktop tool, so the comparison also states where it is the wrong choice.

The problem with most comparisons is a ranking published without its conditions. The same configuration can be excellent for Chinese to English and unusable for Chinese to Vietnamese, accurate in a quiet room and broken the moment background music starts. A ranking without premises is not information.

## Conditions for this comparison

Everything below rests on four premises. Change one and the conclusions need recalculating:

1. **Language pair** — Chinese to English, one direction. This is the strongest tier, so gaps between approaches widen on other pairs;
2. **Audio conditions** — USB condenser microphone, background music present, normal commerce speaking rate. The music level is stated in OBS mixer zones — voice peaking in the yellow band, music held in the green zone below -20 dBFS — so anyone can reproduce the same condition on their own machine;
3. **Latency definition** — speech to the end of the spoken translation, not speech to caption. Those differ by several seconds;
4. **Cost basis** — direct cost per broadcast hour including synthesis, excluding salaries and equipment depreciation.

## The four approaches

**Platform built-ins.** Platform-side captioning and translation roll out by region and app version, and the boundary moves. At the time of writing, what TikTok's [newsroom](https://newsroom.tiktok.com/en-us/auto-translations-and-captions) has announced covers auto-captions and caption/description translation on **videos**; there is no equivalent published specification for real-time translation on LIVE — consistent with what the encoder settings article found, namely that TikTok publishes no citable technical specs. So the honest description of this category is that whether it exists, and how much of it exists, is whatever you see on the day you go live.

Integration cost really is the lowest of the four, with no extra software and no local compute. The price is control: language pairs, terminology, caption styling and segmentation are not configurable, and when something goes wrong there is nowhere to look.

**Desktop tools.** The whole pipeline runs on the host's own machine, feeding OBS through a virtual audio cable. Control is the highest, captions can be laid out bilingually and keyed, and glossaries are yours to maintain. The price is local performance headroom and dependence on a stable connection.

**Self-hosted ASR, MT and TTS pipelines.** You wire the three service categories together yourself. The flexibility ceiling is the highest, and deep customisation for a specific product category becomes possible. The real cost is the glue — segmentation, terminology, audio routing, failure recovery — and that is the most consistently underestimated work in this entire space.

**Live interpreters.** A human interprets in real time. The quality ceiling is the highest, and slang, improvisation and cultural transposition all work, which machines handle poorly. The price is hourly billing and the difficulty of sustaining output across long sessions.

## Side by side

Under the conditions stated above:

| Dimension | Platform built-in | Desktop tool | Self-hosted pipeline | Live interpreter |
| --- | --- | --- | --- | --- |
| Time to first broadcast | Minutes | Half a day to a day | Two to six weeks | Hiring cycle |
| Latency (speech to end of spoken translation) | Not controllable, not published | Tunable via segmentation | Tunable, lowest theoretical floor | Shortest; interpreters anticipate |
| Language pair control | None | Full | Full | Whatever the interpreter has |
| Glossary support | No | Yes | Yes | Verbal briefing |
| OBS caption layout | Not controllable | Controllable | Controllable | Needs a separate captioner |
| Direct cost per hour | Usually zero | Subscription or usage, lowest of the paid options | Lowest API cost, highest engineering cost | Substantially above the other three |
| Stability over long sessions | High | High | Depends on implementation quality | Lowest |
| Who debugs a failure | Nobody | You | You | The interpreter and you |

The table deliberately omits millisecond figures and prices. Both move with platform releases, vendor pricing and your own hardware, so a hard number written here would be wrong within months. Measure latency yourself using the two-track recording method described in the latency budget article.

## Choosing

Working backwards from constraints beats scoring feature lists:

- **Testing whether anyone overseas is watching at all** — use the platform built-in. You can be live in half an hour, and building a pipeline at this stage is premature.
- **Already committed to one language pair long term** — use a desktop tool. This is where control and cost balance out, with captions and terminology under your own management.
- **Stable engineering capacity plus unusual requirements** (an existing terminology database, integration with internal systems, operation inside a private network) — build the pipeline, provided somebody owns it beyond the initial delivery.
- **High order values and infrequent sessions** (large wholesale, B2B recruitment, high-ticket single items) — hire an interpreter. Once one session covers the fee, the quality advantage is not purchasable from the other three.

## Disclosure

This site is maintained by the team behind OBSTrans, known as Yanbo in Chinese. OBSTrans is a **desktop tool** in the categories above, so this is not a neutral third-party review, and it should be read on that basis.

The cases where it is the wrong choice, stated as plainly:

- **Teams testing one or two sessions** do not need it; the platform built-in is less work;
- **Third-tier target languages** such as Indonesian, Vietnamese and Thai cap its translation quality at whatever the upstream model provides, which is no better than a self-hosted pipeline would achieve;
- **Deep customisation**, such as substituting terminology from your own product database, has a higher ceiling with a self-hosted pipeline;
- **Constrained host machines** running a game or several outbound streams alongside the broadcast will contend for resources; moving the pipeline to a second machine or a cloud-side approach makes more sense there;
- **Order values high enough to justify a human** put slang and improvisation beyond any machine approach, including ours.

## When to re-evaluate

The relative positions here are not fixed. Two developments justify recalculating.

The first is **platform built-ins gaining configurability**. The moment platforms expose language pair and terminology settings, they become viable well beyond trial sessions.

The second is **changing tier on your language pair**. Moving from Chinese to English towards South-East Asian languages compresses the gaps, because upstream model quality becomes the binding constraint. Paying for control yields less at that point, and effort spent on scripting and templated replies yields more.


## FAQ

### Is building your own ASR, MT and TTS pipeline cheaper?

The per-call API pricing is genuinely lower than packaged tools, but the saving rarely covers the engineering time. Segmentation strategy, glossaries, audio routing and reconnection logic add up to several weeks of work plus ongoing maintenance. Avoid this route unless someone owns it long term.

### Are platform built-in translation features good enough?

It depends on how much control you need. Integration cost is the lowest available, but language pairs, terminology, caption styling and latency behaviour are not yours to configure, and bilingual caption layout in OBS is not possible. Fine for trial sessions, thin for a primary operation.

### When is a live interpreter worth the cost?

When order values are high and sessions are infrequent. The quality ceiling is the highest of the four, but billing is hourly and sustained output across long broadcasts is difficult. A three-hour daily show rarely survives a month on interpreters.