Speech-to-Text (STT)

STT models available through Telnyx for LiveKit voice agents.

Telnyx hosts Deepgram models on dedicated GPUs. Access them through the Telnyx plugin:

from livekit.plugins import telnyx

stt = telnyx.deepgram.STT(model="nova-3", language="en")

🔌 Plugin source on GitHub

Available models

Nova-3 (recommended)

Latest generation, best accuracy.

stt = telnyx.deepgram.STT(
    model="nova-3",
    language="en",
    interim_results=True,
    keyterm=["YourBrand", "custom-term"],  # keyword boosting
)

Nova-2

Previous generation, stable and reliable. Uses weighted keyword boosting.

stt = telnyx.deepgram.STT(
    model="nova-2",
    language="en",
    interim_results=True,
    keywords=["YourBrand:2.0", "custom-term:1.5"],
)

Flux

Experimental, with built-in end-of-turn detection. Designed for real-time voice agents.

stt = telnyx.deepgram.STT(
    model="flux",
    language="en",
    interim_results=True,
    keyterm=["YourBrand", "custom-term"],
    eot_threshold=0.5,
    eot_timeout_ms=3000,
    eager_eot_threshold=0.3,
)

Parameters

Parameter	Default	Description
model	nova-3	Model to use (nova-3, nova-2, flux)
language	en	Language code
interim_results	True	Stream partial transcriptions
keyterm	—	Keyword boosting (Nova-3, Flux)
keywords	—	Weighted keyword boosting (Nova-2)
eot_threshold	—	End-of-turn confidence threshold (Flux only)
eot_timeout_ms	—	End-of-turn timeout in ms (Flux only)
eager_eot_threshold	—	Eager end-of-turn threshold (Flux only)