Speak and hear the translation

Choose the route for your task

This route uses the dedicated gpt-realtime-translate translation interface. The service detects the source language automatically. Each session produces one target language, using one microphone audio track in this view.

Translated speech can follow the source's general tone, pitch and speaking style. This does not guarantee an exact copy of the speaker's identity or voiceprint. Fixed voices, uploaded reference recordings and custom instructions are unavailable here. See OpenAI's realtime translation guide.

Save the connection once in Setup

1. Open Setup from the top toolbar, go to API configuration, and find OpenAI.
2. Enter your account's API key and Base URL. The official base is https://api.openai.com/v1. Set the realtime model to gpt-realtime-translate.
3. Enable and save the connection, then return to the Studio. Select the saved connection when using it; you do not need to enter its URL or key again.

A compatible service must implement OpenAI's dedicated continuous translation protocol. Support for text chat, file transcription or voice conversation alone is insufficient. Do not enter a dedicated Qwen or Doubao speech address, or append /realtime/translations to the Base URL.

This route sends microphone audio to the selected service and receives translated speech. Check the address, account access and service charges before starting. The app does not switch to another provider automatically.

Start speaking and let the final phrase finish

1. Open Speech Translation Studio and choose Live translation · OpenAI. Wear headphones so that translated speech does not feed back into the microphone.
2. Select a saved OpenAI connection and the Translate into. Source language stays automatic. Enable Also show input transcription if you also want to read the incoming speech.
3. Click Start live translation. Wait for the connection, configuration confirmation and device preparation to finish before speaking. Once the status reads Listening and translating, check the microphone level meter for input activity.
4. Speak at a natural pace, with normal pauses. Audio is processed continuously, and translated text and speech arrive as they become available.
5. Click Stop translation when finished. Capture stops, but captions and audio from the final phrase already sent may still arrive. Wait for the session and final playback to finish before starting again.

Closing the Studio panel leaves interpretation running. Use Stop translation to end capture. Hiding the panel or viewing another part of the app does not end the session.

There is no fixed speech-output latency guarantee. Network conditions, language pair, speaking pace and service load affect the wait. Begin with a short sentence and check the source captions, target language and playback volume before a longer session.

1234

1. 1Choose Live translation to use this OpenAI route.
2. 2Select the saved connection, target language and optional input transcription before starting.
3. 3Start or stop here, and check the microphone level and session state.
4. 4Read each direction separately; Latest returns to incoming text.

Select one target language

The view offers these 13 target languages. Codes below help you check a configuration; select the language name during normal use.

Review mixed-language speech, names, numbers and overlapping speakers. Automatic language detection does not guarantee that every recording will be translated correctly.

Read captions and keep the complete record

Source captions and translated text are appended separately. A word may arrive in several pieces; the app joins them in the service's order. Let a phrase finish before judging the translation from a single fragment.

The source and target reading windows each keep the latest 32 KiB of UTF-8 text. The number of visible characters depends on the language. Earlier text leaves those windows. After stopping the session, use Export record · JSONL to save the complete caption events. Export is unavailable while the session is running. JSONL stores one record per text line and can be opened in a text editor. Copy window copies only the text currently retained. After reading earlier content, use Latest to follow new captions again.

The record contains the target language, caption fragments and session state. It contains no original audio and does not automatically create a WAV file. Records are saved with the app's data directory. Exiting the app cancels active interpretation. Stop normally and let the final phrase finish before exiting.

Turning source captions off removes this source transcription stream from the display. Translated speech and target captions still come from the audio translation. Source transcription itself can contain errors.

12

1. 1After stopping, check that the final phrase remains in the caption window.
2. 2Export record · JSONL saves the full local event record, including text no longer visible in the window.

Check these points if something goes wrong