To generate transcripts with start and end times of spoken sentences by an agent or user in LiveKit, you can use the speech-to-text (STT) plugins provided by LiveKit. These plugins generate

SpeechEvent

objects that include

start_time

and

end_time

properties. Below are some examples: • *Baseten Plugin:*`event = stt.SpeechEvent( type=stt.SpeechEventType.FINAL_TRANSCRIPT, alternatives=[ stt.SpeechData( language=language, text=text, confidence=confidence, start_time=start_time, end_time=end_time, ) ], )` • *Cartesia Plugin:*`speech_data = stt.SpeechData( language=language, start_time=0, # Cartesia doesn’t provide word-level timestamps in this version end_time=data.get("duration", 0), # This is the duration transcribed so far confidence=data.get("probability", 1.0), text=text, )` • *Sarvam Plugin:*`# Try to get timestamps if available timestamps_data = response_json.get("timestamps") if timestamps_data and isinstance(timestamps_data, dict): words_ts_start = timestamps_data.get("start_time_seconds") words_ts_end = timestamps_data.get("end_time_seconds") if isinstance(words_ts_start, list) and len(words_ts_start) > 0: start_time = words_ts_start[0] if isinstance(words_ts_end, list) and len(words_ts_end) > 0: end_time = words_ts_end[-1]` Additionally, the

TranscriptSynchronizer

component can handle timed text through the

push_text

method, which accepts text with timing annotations:

def push_text(self, text: str) -> None:
    start_time, end_time = None, None
    if isinstance(text, io.TimedString):
        start_time = text.start_time or None
        end_time = text.end_time or None

You can use these examples to implement transcript generation with timing information in your application. Sources: GitHub::livekit::agents | livekit-plugins/livekit-plugins-sarvam/livekit/pl...