How to Translate Speech in Real Time

How real-time translation works

Speech recognition first produces source text, then the selected local LLM translates it. Normal mode translates a finalized segment. Low Latency mode also sends the changing prediction to the translation task, so it does not have to wait for the entire spoken segment.

Set up Ollama and the translation task

Recommended real-time translation model

Hy-MT2-1.8B

On-device and real time

A compact model from the Hunyuan Translation 2.0 family, focused on translation across 33 languages plus five Chinese minority-language or dialect varieties. It is well suited to local, real-time translation in Owl Meeting.

• The publisher reports leading results on the open FLORES-200 and WMT25 benchmarks, outperforming several mainstream commercial translation APIs.
• Supports structured, delimiter-aware, contextual, glossary-guided, and style-controlled translation instructions.
• Only 1.8B parameters; its smallest 1.25-bit build is about 440 MB, balancing quality, speed, and on-device resource use.

API model name

hunyuan-mt2-1.8b-chat

Maximum input

4K

Maximum output

4K

View the model on Ollama →

1. Install and start Ollama, then download a model that supports the languages you need.
2. On the AI page, create a Translation task and set its model, source language, target language, and name.
3. Set Input mode to Per segment. Batch and Full text modes are intended for saved transcripts, not live translation.

Choose the right recognition mode

Normal mode

Provides stable recognition results. Translation starts after the current speech segment is finalized.

Low Latency mode

Shows predicted text earlier and translates those predictions in real time. Use this mode when reading the translation during a meeting, lecture, stream, or video.

Start live translation

1. Choose Microphone for local speech, System audio for video or remote participants, or Dual input for both.
2. Select a speech-recognition model that supports the language being spoken.
3. Select the enabled translation task and click its blue start button to load the local LLM.
4. Start recording. Normal mode adds the translation after each finalized segment; Low Latency mode updates it while speech is still in progress.

Reduce translation delay

• Enable Low Latency mode when you need translation before the segment ends.
• Use a smaller, faster translation model and ask it to output only the translation.
• Reduce other CPU, memory, or GPU workloads and verify that the recognized source text is accurate.

Record the setup before comparing latency

Translation delay is not determined by Owl Meeting alone. Recognition mode, the speech model, the translation model and quantization, prompt length, CPU or GPU, available memory, and competing workloads all affect when text appears.

• Test a short sample that contains the real source and target languages.
• Record the Owl Meeting version, model names, quantization, hardware, and audio source.
• Compare time to the first useful translation and the number of revisions, not only total model speed.

Verify each language pair and important term

A multilingual model may not provide equal quality in every direction. Test names, product terms, numbers, and sentence boundaries on the language pairs that matter to the meeting. A glossary or explicit task instruction can improve consistency, but it cannot repair an incorrect source transcript.

• Check the model card instead of assuming every LLM supports every language.
• Review the recognized source text before diagnosing a translation problem.
• Keep the requested output concise so partial results remain readable in Low Latency mode.

Understand privacy and live-result limits

With a local recognition model and a local Ollama address, recognition and translation can run on the computer. If the task points to an external service, its data handling rules apply instead; verify the configured endpoint before processing a sensitive meeting.

Low Latency mode translates changing predictions. Early words can be revised as more speech arrives, and noise or overlapping speakers can affect both the source and translation. Treat the live result as assistance and review important wording before sharing or exporting it.

Frequently asked questions

Test the full path with a short sample

Check the audio source, recognition language, translation direction, selected LLM, and latency before a long meeting.

Get Owl Meeting for WindowsOpen the AI assistant documentation