To address the issue of multiple transcripts being sent for a single user utterance, you can configure Deepgram STT to send only truly final transcripts by disabling interim results. Here's how you can do this: Python Configuration:

from livekit.plugins.deepgram import STT

deepgram_stt = STT(
    model="nova-3",
    interim_results=False,  # Disable interim results
    endpointing_ms=25,      # Control silence detection
    punctuate=True,
    smart_format=True,
    language="en-US"
)

JavaScript Configuration:

import { STT } from '@livekit/agents-plugin-deepgram';

const deepgramSTT = new STT({
    interimResults: false,  // Disable interim results
    endpointing: 25,        // Control silence detection
    punctuate: true,
    smartFormat: true,
    language: 'en-US'
});

By setting

interim_results

(Python) or

interimResults

(JavaScript) to

false

, the system will only emit final transcripts and skip interim ones. Additionally, you can configure the

endpointing_ms

parameter to control when speech is considered ended. For example: •

endpointing_ms

: Time in milliseconds of silence to consider end of speech. Set to 0 to disable. Defaults to 25. •

punctuate

: Whether to add punctuations to the transcription. •

smart_format

: Whether to apply smart formatting to numbers, dates, etc. Regarding the TranscriptSynchronizer, I couldn't find specific details about its ability to buffer or deduplicate transcripts. If you need further assistance with this component or using the

conversation_item_added

event to intercept and control transcript forwarding, please let me know, and I can escalate this to a specialist for more detailed guidance. Sources: livekit-plugins/livekit-plugins-aws/livekit/plugins/aws/stt.py | plugins/deepgram/src/stt.ts