Speech translation is the task of converting spoken audio in one language into text, or sometimes speech, in another. It combines recognition and translation, either as two separate steps or in a single end-to-end model, and it is a common feature layered onto speech-to-text APIs for multilingual products.
How speech translation works
Speech translation uses one of two architectures. The cascade approach runs transcription in the source language, then feeds that text to a machine-translation model for the target language, optionally followed by text-to-speech.
The end-to-end approach uses one model that maps source audio directly to target-language text; the Whisper model, for instance, can translate speech into English in a single pass.
The cascade is more flexible and often more accurate across many language pairs, but its errors compound across stages, since a recognition mistake becomes a translation mistake. The end-to-end model is simpler and lower-latency, but usually more limited in the directions it supports.
Why it matters for speech-to-text
Many providers advertise speech translation as a speech-to-text add-on, and the two architectures behave differently enough that you should know which one you are buying. A cascade lets you swap or tune the translation stage and typically covers more language pairs, at the cost of one recognition error turning into a translation error.
An end-to-end model is easier to integrate but is frequently restricted to translating into English only.
Where speech translation is used
Live translation of meetings, webinars, and support calls is the headline use case, where a caller speaks one language and the other side reads or hears another in near real time.
Post-hoc translation is just as common: subtitling foreign-language video, making recorded interviews searchable in English, or standardizing a multilingual archive into one working language for review. The real-time versions lean on the same low-latency streaming and language identification that live multilingual ASR depends on.
Practical notes on speech translation
With speech translation, check the supported language pairs and the direction carefully, because a surprising number of systems translate only into English. Expect lower accuracy than same-language transcription, since translation adds its own errors on top of recognition.
Also confirm what survives the process: timestamps, speaker diarization, and formatting often do not carry cleanly from the source audio into a translated transcript.
Frequently asked questions
What is speech translation? Turning spoken audio in one language into text or speech in another, by pairing speech recognition with machine translation, either as two stages or in one model.
How does speech-to-text translation work? Either a cascade, which transcribes then translates the text, or an end-to-end model that maps source audio straight to the target language.
Which systems offer two-way translation? Fewer than you would expect. Many translate only into English, so verify both directions and the exact language pairs before you commit.