Overview
This plugin allows you to use Nabrah as an STT provider for your voice agents. Nabrah is a streaming speech-to-text service for Arabic, and the plugin includes end-of-turn detection tuned for Arabic conversation.
Installation
Install the plugin:
uv add "livekit-agents[nabrah]~=1.8"
Authentication
The Nabrah plugin requires a Nabrah API key.
Set NABRAH_API_KEY in your .env file:
NABRAH_API_KEY=<your-nabrah-api-key>
Usage
Use Nabrah STT in an AgentSession or as a standalone transcription service. For example, you can use this STT in the Voice AI quickstart.
The default eot_nabrah model emits an end-of-turn signal, so let the STT drive turn detection:
from livekit.agents import AgentSessionfrom livekit.plugins import nabrahsession = AgentSession(stt=nabrah.STT(recognition_model="eot_nabrah",language="ar-SA",),turn_handling={"turn_detection": "stt", "endpointing": {"min_delay": 0}},# ... llm, tts, etc.)
Turn detection
With recognition_model="eot_nabrah", the recognizer signals when the speaker has likely finished a turn. The plugin waits end_of_turn_confirm_delay_seconds (0.4 seconds by default) after that signal before committing the turn. If the user speaks again inside that window, the turn stays open. Set it to None to commit as soon as the signal arrives.
Nabrah's default model (recognition_model="") is more accurate but doesn't emit an end-of-turn signal. When you use it, configure another turn detector, or set max_transcript_inactivity_seconds as a fallback that commits the turn after that many seconds without new transcript text. The fallback is off by default because it measures recognizer inactivity, not acoustic silence, so a provider delay can split continuous speech.
To learn more, see Turn detection.
Word boosting
Bias recognition toward names and terms the model gets wrong with priority_words. Multi-word phrases are boosted as a unit:
stt = nabrah.STT(priority_words=["مستشفى الملك فيصل التخصصي", "رقم الهوية الوطنية"],priority_words_strength=0.5,)
0.5 is the recommended strength. Higher values make boosted terms appear where they weren't said, so keep the list to terms your users actually say.
For longer lists, keep the terms in a JSON file so you can edit them without touching code. Read the file with encoding="utf-8" so Arabic terms load correctly on every platform:
import jsonimport pathlibboosting = json.loads(pathlib.Path(__file__).with_name("boosting.json").read_text(encoding="utf-8"))stt = nabrah.STT(priority_words=boosting["words"],priority_words_strength=boosting["boost_threshold"],)
Parameters
This section describes some of the available parameters. See the plugin reference for a complete list.
recognition_modelstringDefault: eot_nabrah"eot_nabrah" emits the end-of-turn signal used for turn detection. "" selects Nabrah's default model, which is more accurate but emits no signal. See Turn detection.
languagestringDefault: ar-SARecognition language.
end_of_turn_confirm_delay_secondsfloatDefault: 0.4Seconds to wait after an end-of-turn signal before committing the turn. None commits immediately.
max_transcript_inactivity_secondsfloatDefault: NoneCommits the turn after this many seconds without new transcript text. Disabled by default.
priority_wordslist[string]Default: []Words and phrases to bias recognition toward. See Word boosting.
priority_words_strengthfloatDefault: 0.5How strongly to bias recognition toward priority_words.
disable_number_normalizationbooleanDefault: falseSet to True to turn off Nabrah's number normalization.
api_keystringDefault: env:NABRAH_API_KEYNabrah API key. Defaults to the NABRAH_API_KEY environment variable.
Additional resources
The following resources provide more information about using Nabrah with LiveKit Agents.
