Messages

Messages you receive

Metadata

On connect, with the session’s request_id.

Results

The text of the current segment, in channel.alternatives[0].transcript, always whole words. An interim, is_final: false, replaces the one before it. A final, is_final: true, holds everything since the previous final, so joining the finals with one space gives the transcript.

SpeechStarted

A segment has begun.

UtteranceEnd

Follows the final that ends a segment, with last_word_end.

When text becomes final

End of a segment

After the pause set by utterance_end_ms: a final with speech_final: true, then UtteranceEnd.

Finalize

At once: a final with from_finalize: true.

Long speech

A segment still open after 30 seconds is finalized at its last word boundary, with speech_final: false.

CloseStream

What is still open goes out as a final, then the socket closes.

Words

Each entry of words is one word, with word, punctuated_word, start and end in seconds, and confidence, always 1.0. In Chinese and Japanese, which are written without spaces, each character is an entry and the transcript gains no space. When you name the language, languages repeats it.

Messages you send

Audio

Binary messages, in chunks of any size, at the speed of speech.

{"type": "KeepAlive"}

Accepted. It does not keep a quiet session open: a session closes after 30 seconds without audio, so send silence instead.

{"type": "Finalize"}

Finalizes the current segment now.

{"type": "CloseStream"}

Sends the last final, then closes the socket.

Errors and close codes

A parameter the socket cannot honour is refused with an error that names it. When a session ends on an error, an Error message with the reason comes first. The socket then closes with one of these codes: