Speechmatics STT plugin for LiveKit Agents

Support for Speechmatics STT.

See https://docs.livekit.io/agents/integrations/stt/speechmatics/ for more information.

Installation

pip install livekit-plugins-speechmatics

Diarization

Speechmatics STT engine can be configured to emit information about individual speakers in a conversation. This needs to be enabled using enable_diarization=True. The text output of the transcription can be configured to include this information using the macros speaker_id and text, as shown in the examples below.

<{speaker_id}>{text}</{speaker_id}> -> <S1>Hello</S1>
[Speaker {speaker_id}] {text} -> [Speaker S1] Hello

You should adjust your system instructions to inform the LLM of this format for speaker identification.

Usage (Speechmatics end of utterance detection and speaker ID)

To use the Speechmatics end of utterance detection and speaker ID, you can use the following configuration.

Note: The turn_detection_mode parameter tells the plugin to control the end of turn detection. The default mode is ADAPTIVE, which means that the plugin will control the end of turn detection using the plugin's own VAD detection and the pace of speech. In the example below, we use the default ADAPTIVE mode. The turn_detection="stt" parameter tells the plugin to use the STT engine's end of turn detection.

from livekit.agents import AgentSession
from livekit.plugins import speechmatics

agent = AgentSession(
    stt=speechmatics.STT(
        speaker_active_format="[Speaker {speaker_id}] {text}",
        speaker_passive_format="[Speaker {speaker_id} *PASSIVE*] {text}",
        additional_vocab=[
            speechmatics.AdditionalVocabEntry(
                content="LiveKit",
                sounds_like=["live kit"],
            ),
        ],
    ),
    turn_detection="stt",
    ...
)

Usage (LiveKit Turn Detection)

To use the LiveKit end of turn detection, the format for the output text needs to be adjusted to not include any extra content at the end of the utterance. Using [Speaker S1] ... as the speaker_active_format should work well. You may need to adjust your system instructions to inform the LLM of this format for speaker identification. You must also include the listener for when the VAD has detected the end of speech.

The end_of_utterance_silence_trigger parameter controls the amount of silence before the end of turn detection is triggered. The default is 0.5 seconds.

Usage:

from livekit.agents import AgentSession
from livekit.plugins.turn_detector.multilingual import MultilingualModel
from livekit.plugins import speechmatics, silero

agent = AgentSession(
    stt=speechmatics.STT(
        end_of_utterance_silence_trigger=0.2,
        speaker_active_format="[Speaker {speaker_id}] {text}",
        speaker_passive_format="[Speaker {speaker_id} *PASSIVE*] {text}",
    ),
    vad=silero.VAD.load(),
    turn_detection=MultilingualModel(),
    min_endpointing_delay=0.3,
    max_endpointing_delay=5.0,
    ...
)