If you search through every setting during a live stream and still cannot find a 'real-time translated subtitles' switch, most of the time the problem is not a software fault; it is that subtitles in a live stream are not a single button. They are a pipeline: audio capture sends the voice into the system, speech recognition turns the sound into text, the translation service converts the text into the target language, and the subtitle layer overlays the translated text onto the streamed video. If any of the four stages is not connected, the result on the viewer's side is the same—no subtitles. Below, we explain the methods, solution selection, and troubleshooting along this pipeline.
First, Clarify: Do You Want 'Subtitles' or 'Translated Subtitles'?
Before making changes, first check which layer is currently missing on the viewer's side; this determines which entry point you need to modify.
- No Subtitles at All: Capture or recognition is not passing; first confirm whether the recognition side is returning the original text, and do not rush to switch translation services.
- Source-Language Subtitles but No Translation: Recognition is already working; the problem lies in the translation step, or the subtitle layer is only bound to the original-text output.
- There is a translation, but it displays incorrectly:issues with the language pair, sentence segmentation, styling, or occlusion; the pipeline itself is working.
The fixes for these three cases are completely different. Diagnose clearly before acting, and you can avoid most ineffective trial and error.
The four-stage pipeline: where the audio comes from and where the translation appears on screen
- Audio capture.The input is the host's microphone and may also include co-streaming guests and audio played by the system; the output should be a clean, voice-focused audio stream that does not clip. If you send a lot of background music to the recognition stage, you get lyrics instead of spoken lines.
- Speech recognition.It takes an audio stream as input and outputs text segments with timing information. Sentence segmentation, punctuation, numbers, and names are finalized here, and this is also where the ceiling for translation quality is set.
- Translation.It takes recognized text as input and outputs sentences in the target language. The language pair direction, glossary, and whether whole sentences are submitted determine whether the translation is readable.
- Subtitle composition and on-screen display.Lay out the translated text in one or two lines and overlay it via the live streaming software's text source or browser source, or use the platform's built-in subtitle channel. Font size, position, one- or two-line layout, and refresh method are all set at this step.

How to choose among the three implementation paths: platform built-in, live streaming software plug-ins, and an external subtitle layer
| Path | Use cases | Degree of control | Main limitations |
|---|---|---|---|
| Platform built-in subtitles and translation | The platform already provides this capability, and the language pairs cover your audience. | Low; usually only on/off, font size, and position. | Limited by the languages and regions supported by the platform; terminology and styling generally cannot be changed. |
| Live streaming software plug-ins or extensions | Live streaming with desktop streaming software requires you to control the visuals and subtitle styling yourself | Medium-high, with optional recognition and translation services and adjustable sentence segmentation | Requires setting up audio routing yourself, and plugin quality and maintenance status vary widely |
| External subtitle layer | Simultaneous multi-language, multi-platform streaming, with many terms or a need for branded styling | Highest, as every segment can be replaced | The entire pipeline must be maintained by you, and when issues arise you have to troubleshoot segment by segment |
There are really only three criteria: whether the platform has this capability, whether your language pair is covered, and whether you need to control the visual styling yourself. If the first two are met, start with the built-in option; only when neither is met is it worth building an external subtitle layer. The higher the degree of control, the more parts you need to maintain.
Using an OBS external subtitle layer as an example: a minimal viable setup
The following setup does not depend on a specific plugin; any streaming software that supports browser sources and text sources can follow it.
- Separate the audio.The microphone should be on its own channel, with system audio and guest audio each on separate channels. The recognition side should take only the vocal track; music and sound effects should not enter recognition.
- Connect speech recognition.Choose a recognition service or local model that supports streaming responses, verify that the sample rate, channel configuration, and input device match, and use a short test recording to confirm it returns the source text.
- Send the recognition results to translation.Submit by sentence rather than word by word, enable punctuation and sentence segmentation, and build a glossary first for high-frequency terms such as product names, model numbers, and sizes.
- Generate subtitle text.Output one line of source text followed by one line of translation, set a character limit for each line—about 15 to 20 Chinese characters and about 40 characters for English—and truncate or wrap when the limit is exceeded.
- Overlay it onto the video.In the streaming software, add a browser source, align its width with the streaming resolution, keep only the subtitle area in its height, and place it within the bottom safe area of the frame; also keep a text source as a fallback.
- Keep the original video available as a fallback.Put subtitles in a separate scene or source group so that if an issue occurs, they can be hidden with one click without affecting the continued broadcast of the main video.

Latency, sentence segmentation, and mistranslation: troubleshoot in this order.
If subtitles do not appear, appear slowly, or appear in a jumbled way, check from upstream to downstream without skipping steps.
- Check whether audio reaches the recognition service.Check whether the recognition service returns the source text. If there is no source text, the issue is in capture or recognition, not translation.
- The source text exists, but sentence segmentation is messy.Check the silence threshold, sentence-break wait time, and whether punctuation is enabled. If one sentence is split into three segments, the translation will inevitably be fragmented.
- The source text is normal, but the translation is wrong.Check whether the language pair direction is reversed, whether the glossary is taking effect, and whether word-for-word translation is being used.
- The translation is correct, but viewers cannot see it.Check subtitle layer refresh, browser source layering, whether it is obscured by other assets, and whether streaming buffering is normal.
- Everything is normal, but the overall process is slow.Latency mainly comes from waiting for a sentence to finish and from translation request round trips. Shortening the sentence segmentation window and switching to on-device recognition can reduce part of it, but it cannot be eliminated completely.
Run an end-to-end acceptance check from the viewer's perspective before going live
Checking the viewer side on a phone reveals more problems than previewing on the streaming side. Before going live, go through each item:
- On the viewer side, either only translated subtitles are displayed or both languages are complete, and they are not obscured by platform UI elements.
- Say a sentence with proper nouns, and confirm that terms are translated according to the glossary rather than paraphrased into other words.
- Mute for ten seconds, then speak; the subtitles should recover and not get stuck on the previous sentence.
- After switching scenes or inserting a product image, subtitles still appear on the new image.
- Disconnect and reconnect the network once; the subtitle layer should reconnect and there should be no old sentences left on the screen.
Subtitles solve the understanding problem; comments and direct messages still need to be handled
Live subtitles only address the understanding problem during the broadcast period. After the stream ends, viewers will move to comments, direct messages, and communities to continue asking about price, size, logistics, and after-sales. These messages are scattered across multiple social platform accounts and in different languages; it is difficult for a streamer alone to keep up by switching accounts and languages.
SKY SCRM is designed for exactly this stage: it supports multi-account management across global social platforms, multilingual communication, AI translation, and customer management, and is suitable for handling cross-language inquiries and follow-up brought by live streams. It does not handle subtitle generation and on-screen display within the live stream; the two are different parts of the chain and should not be confused. The instant translation entry on mobile is also not the same mechanism as the live subtitle layer; for the difference, seeHow to enable real-time translation on mobile; if the subtitle issue occurs in a meeting scenario rather than a live stream,Video Conferencing Real-Time Translation Subtitle Settingsthe pre-meeting check approach in them also applies; live streaming simply adds two more parts: stream delivery and viewer-side display.
Frequently asked questions
Can I do live real-time translated subtitles without installing OBS?
Yes, provided that the platform you stream on has built-in subtitle or translation capabilities. In this case, subtitles are generated by the platform, and you can only control the on/off switch and a few style options; if the platform does not have this capability, you need streaming software or a third-party tool that can overlay an external subtitle layer.
Can I choose separate services for speech recognition and translation?
Yes. The advantage of choosing them separately is that you can select translation based on language pair and terminology needs, and select recognition based on audio quality; the cost is that you need to integrate the two stages yourself, sentence segmentation and submission methods need to be debugged, and when errors occur you also need to troubleshoot each stage separately.
Why are subtitles always two or three seconds late?
This mainly comes from the cumulative effect of three stages: waiting for a sentence to finish, recognition returning, and the round trip of the translation request. Shortening the segmentation wait can significantly improve it, but waiting too short will chop sentences into pieces and make the translation harder to read; you need to find a balance at the actual speaking rate.
Will bilingual subtitles block the view?
Two-line subtitles are taller than single-line subtitles, especially on small phone screens. Restricting subtitles to the bottom safe area, limiting characters per line, and adding a semi-transparent background bar behind the text usually balances readability and the picture. Product information and the area where people speak to camera should be avoided in advance.
Do viewers need to install a plugin to see the subtitles?
No. The subtitles are overlaid on the streamed image, so viewers see the same image you send out. For this reason, once the subtitle position covers key information, all viewers will be affected.

