Skip to main content
Word timestamps tell you when each word was spoken, so you can highlight words as they’re said, line up captions, redact audio, or measure how quickly your agent responds. ink-preview includes them in every transcript, with nothing to turn on.

Which models return them

What each word looks like

Each entry in the words list has three fields: Times are measured from the first audio you send on the connection, not from the start of each turn. They’re rounded to the nearest 10 ms.

Realtime STT (Auto)

turn.update, turn.eager_end, and turn.end carry words, just as they carry transcript. turn.update has the words so far, and turn.eager_end and turn.end have every word in the turn. Each new turn starts a fresh list. A word is added once the model has heard all of it, and its entry never changes after that. So while someone is still talking, words can trail the transcript slightly. Here, someone is in the transcript but not in words yet:
words is left out until the turn has its first complete word, so treat a missing words as an empty list.

Realtime STT (Manual)

On Realtime STT (Manual), each transcript event includes only the words that finished since the last event, just as text includes only the new text. Keep a running list and add each event’s words to it. When flush_done or done arrives, you have every word for the audio you finalized. Timestamps can trail the text slightly. In this example, away is in the first event’s text, but its entry arrives with the second event. satellite is in the second event’s text, and its entry arrives later still:
Late words can also arrive in an event with empty text. Keep the words, and add nothing to the transcript.

How words are split

In languages written with spaces, each entry is one word. Punctuation goes with the word before it, as in away, and please., but opening quotes and brackets go with the word after, as in "Hello and (see. In languages written without spaces, such as Chinese, Japanese, and Thai, the model decides where to split, so an entry can be part of a word, a whole word, or a few words. English words and numbers inside Chinese or Japanese text come back as their own entries. Either way, once all the words have arrived, the entries cover the whole transcript, in the correct order.

How accurate the times are

The model predicts each timestamp rather than measuring it, so expect them to be close but not always exact:
  • A long word, such as a spelled-out email address or a long number, can get a start that’s later than when the speaker began the word.
  • An entry can have zero length, with start equal to end. This is rare in languages written with spaces, and common in Mandarin, Cantonese, Japanese, and Thai, where a word can come back in pieces. Its text still belongs in the transcript.

Where to go next

Turn detection

How words fits into the turn lifecycle

Languages

Detected languages arrive on the same events

API reference

Every event and field