Boundless FlowSpeech workstation · User manual
0.5.0简体中文
USER GUIDE / 05

Multilingual translation

Test text translation first, recorded subtitles second, and interpretation last.

Screenshots show the app interface with example settings and sessions. Use your own models, services and recordings.

First time here? Start with a taskFollow illustrated tutorials from installation and model setup to your first result →

Select a translation service

  1. Open API Config in the setup wizard, find the required translation or LLM service, and enter a model ID actually served by the endpoint.
  2. Select it in Translation service, enter source text and target language, and run the translation test. The test does not save setup or create a meeting.
  3. Finish setup to save, then enable translation in recording or interpretation settings and select the primary target plus additional subtitle languages.
Setup guide · Select and test translation Example
  1. 1Select a configured translation service.
  2. 2Enter a short passage to test.
  3. 3Choose the target language.
  4. 4Run the test after entering valid settings.

HY-MT2

HY-MT2 supports a compatible API and a managed service using downloaded local files. Open its card, select the matching runtime and complete GGUF file, then choose Compute device. For NVIDIA GPU, leave CUDA component empty for automatic discovery, or choose the matching ggml-cuda.dll if it is not found. The component and driver must be compatible; select CPU if the device check fails. Test a short translation after starting, then click Use for translation and finish setup to save.

The local Index-Translate form uses the same device options. Restoring a saved configuration verifies the CUDA component again. After replacing a component, restart it and click Use for translation again. An original training checkpoint and a GGUF service package are different inputs; renaming the file extension is not conversion. Loading and the first inference can take longer, so complete a short text test before live interpretation.

HY-MT2 · Endpoint, model and local service settings Example
  1. 1Select NVIDIA GPU, or keep CPU.
  2. 2Leave the CUDA field empty for automatic discovery, or choose a file.
Local translation · Compute device and optional CUDA component (service not started) Example
  1. 1Select NVIDIA GPU, or keep CPU.
  2. 2Leave the CUDA field empty for automatic discovery, or choose a file.

Index-Translate

Use the Index-Translate card for its public API or your own compatible local service, or start a managed service from downloaded files. Public endpoint availability, model names and limits are controlled by the service. Use the current card and actual response as the reference. A successful save does not confirm service availability.

Subtitles and recovery

Chinese speech, English subtitles

  1. Set the recording source language to Chinese / zh and the translation target to English / en. The source controls recognition; the target controls translation. Keep recognition set to Chinese for Chinese speech.
  2. Check the source transcript against the recording first. If it is wrong, check the microphone, recognition model and source language. Translation cannot reliably fix recognition errors.
  3. For consistent technical terms, names or product names, enter a glossary for the relevant target language. Translation uses limited earlier context to resolve ambiguity, while translating only the current segment.
  4. Check that negation, numbers, conditions and questions are preserved. Complete sentences can be translated first. An unfinished price, number or condition may wait for the following words; the model should not invent a conclusion. Held text is processed when the recording segment ends.

If a short utterance has no translation yet, check the source transcript and whether recording is still in progress, then wait for the next words or the segment to end. Subtitle updates balance timeliness with preserving the full meaning.

The setup text test shows the current source and target languages, using the source selected in Models & languages. Enter text in that language and test again after changing it; edited text is kept. If the response clearly conflicts with the target language, the test discards it without retrying automatically. Check the settings and try again manually. This check cannot detect every language or meaning error.

Normal translations display without waiting for a separate proofreading pass. A clearly incompatible writing system can trigger one correction attempt, adding delay and possibly a cloud request charge for that segment. If it still fails, use Translation Recovery. This check cannot detect every language or meaning error.

The current ONNX/FunASR live pipeline supports up to three translations sharing one API, with a separate request per language. The primary language drives speech, floating subtitles and typing; changes apply to the next recording. Additional languages add requests and time. Verify one target before adding more.

For a failed segment, inspect Translation Recovery and explicitly retry. Retrying cloud requests may incur charges again; inspect translation separately from recording. Meeting export can include saved target languages, but does not translate missing segments. Complete those through recovery first.

Live subtitles · Source and translated text Example
  1. 1Interim recognition appears here.
  2. 2Translations are grouped by language and segment.
Meeting export · Language and file format Example
  1. 1Choose the original or a saved translation.
  2. 2Choose a subtitle or text format.
  3. 3Export and choose a save location.