Streaming transcription processes an ongoing audio stream and emits partial text before the speaker has finished. It maintains enough state to connect new audio with earlier context and may revise uncertain words as more evidence arrives.
Latency is a central design choice. Short delays make captions and voice interfaces feel responsive, while longer delays give the system more context for punctuation, ambiguous words and sentence boundaries.
Some applications use a fast stream for immediate feedback and a slower stream for corrections. This lets users see useful text quickly without treating the earliest hypothesis as the final transcript.
ELI5
Streaming transcription writes down speech a little at a time while someone is still talking. It does not wait for the whole recording to end.
For example, live subtitles may show a word immediately and then fix it when the next few words make the sentence clearer. The system trades a small delay for a better guess.
