Skip to content

How live speech translation works: Bringing another language to radio

Listen to Story · 6 min listen
Quick Takeaway · Key Summary

A musician is discussing a new album in Turkish. A listener in another country begins hearing the conversation in English while it is still underway. That is the possibility live speech translation opens: a single interview can become accessible across languages before the conversation is over.

The journey from studio microphone to listener is not instantaneous. A system receives the sound, processes it and produces text or speech in another language. If it mishears a concert date or fails to recognise an artist’s name, the mistake can travel into a fluent-sounding translation.

The journey of a sentence

One approach turns speech into text, translates that text and, if spoken output is needed, voices the translation. Each stage supplies the next. A name misheard at the beginning can appear in the transcript and then be repeated by the synthetic voice.

Not every system takes that route. Google Research’s 2025 explanation compares staged approaches with research into translating audio directly into speech in another language. That is a difference in design; it does not alone tell a broadcaster which product will work best in a particular studio.

Google describes a two-second internal lookahead setting for its model and says inference time also contributes to total latency. Two seconds is therefore not a promise that a radio listener will hear a complete result exactly two seconds later. The examples cover English in both directions with Spanish, German, French, Italian and Portuguese; they do not establish performance in Turkish.

Why can captions change as you read?

When a speaker begins a sentence, its ending is not yet known. A system that waits gains context; one that displays an early translation may revise it as more words arrive. The audience can see something sooner, but changing text may be harder to follow.

A 2020 study by Google researchers examined this relationship between delay and changing intermediate captions. The approach was developed on English-to-German speech and its final evaluation extended to other target languages. Those findings do not establish identical results for another product or a Turkish broadcast.

📻
RadioTEDU Live Stream

Listen while exploring this story

Tune into campus radio with non-stop independent music and culture.

Is a prepared clip the same as a live stream?

Feeding a system a recording already divided into tidy sentences is a different condition from processing a continuous interview with pauses and unexpected transitions. A 2025 literature review by Sara Papi and colleagues notes that much simultaneous speech-to-text translation research has used speech segmented by people in advance.

For a radio team, a sentence translated successfully in a short prepared demo does not establish how a longer continuous interview will perform. A useful trial would include the pauses and speaker changes expected in the actual programme. The review concerns speech-to-text translation; it does not measure a particular speech-to-speech product.

The small details in a radio conversation

Imagine a guest gives a concert date while music plays underneath. If the system mishears the day, the translation may sound polished while giving the audience the wrong information. A missing negative can reverse an answer, and overlapping speakers can make it difficult to tell who said what.

Speed alone is a poor measure of success. A bilingual editor can check names, numbers and meaning against the original audio. Results from a Turkish-to-English test should not automatically be assumed to hold in the other direction.

When translation stops

A broadcasting team also needs to think about the moments when a tool fails. Returning to the original audio or using a prepared explanation can help listeners understand what is happening. The time to the first intelligible translation and the time until final captions stop changing are different observations. In a rehearsal recording, for example, note when the speaker begins a sentence, when an intelligible translation is first heard and when the captions change for the last time. Comparing those points in the same recording makes clear which delay you mean; this is a suggested exercise, not a completed measurement.

If audio is sent to an external service, discuss the scope of that use with participants and review the provider’s retention terms. In a media production setting such as RadioTEDU Technology, the question is not only how quickly another voice appears. It is whether a listener can follow the right person, the right information and the meaning of the conversation.

RadioTEDU monthly

The month in podcasts.

Receive podcast episodes from the latest 30 days and upcoming RadioTEDU ticketed events. No daily mail or unrelated announcements.

This consent covers only the monthly podcast newsletter and ticketed event updates. See our privacy notice.

Live RadioTEDU Pick a station and start listening
0:00 0:00

RadioTEDU · Big5 Türkiye Watch Party

RadioTEDU Live Screening: Big5 Türkiye Watch Party. Thursday 19 November, 18.00, D032. Click for free tickets.