Namo Turn Detector

The Namo Turn Detector v1 utilizes a custom fine-tuned model from VideoSDK to accurately determine whether a user has finished speaking. This allows for precise management of conversation flow, especially in cascading pipeline setups. It can operate as a multilingual model or be configured for a specific language for optimized performance.

Installation

Install the Turn Detector-enabled VideoSDK Agents package:

pip install "videosdk-plugins-turn-detector"

Importing

from videosdk.plugins.turn_detector import NamoTurnDetectorV1

Example Usage

1. For a specific language (e.g., English):

from videosdk.plugins.turn_detector import NamoTurnDetectorV1, pre_download_namo_turn_v1_model
from videosdk.agents import CascadingPipeline

# Pre-download the English model to avoid delays
pre_download_namo_turn_v1_model(language="en")

# Initialize the Turn Detector for English
turn_detector = NamoTurnDetectorV1(
  language="en",
  threshold=0.7
)

# Add the Turn Detector to a cascading pipeline
pipeline = CascadingPipeline(turn_detector=turn_detector)

2. For multilingual support:

If you don't specify a language, the detector will default to the multilingual model, which can handle various languages.

from videosdk.plugins.turn_detector import NamoTurnDetectorV1, pre_download_namo_turn_v1_model
from videosdk.agents import CascadingPipeline

# Pre-download the multilingual model
pre_download_namo_turn_v1_model()

# Initialize the multilingual Turn Detector
turn_detector = NamoTurnDetectorV1(
    threshold=0.7
)

# Add the Turn Detector to a cascading pipeline
pipeline = CascadingPipeline(turn_detector=turn_detector)

Configuration Options

• language: (Optional, str): Specifies the language for the turn detection model. If left as None (the default), it loads a multilingual model capable of handling all supported languages.

• threshold: (float) Confidence threshold for turn completion detection (0.0 to 1.0, default: 0.7)

Supported Languages

The NamoTurnDetectorV1 supports a wide range of languages when you specify the corresponding language code. If no language is specified, the multilingual model will be used.

Here is a list of the supported languages and their codes:

Language	Code
Arabic	ar
Bengali	bn
Chinese	zh
Danish	da
Dutch	nl
English	en
Finnish	fi
French	fr
German	de
Hindi	hi
Indonesian	id
Italian	it
Japanese	ja
Korean	ko
Marathi	mr
Norwegian	no
Polish	pl
Portuguese	pt
Russian	ru
Spanish	es
Turkish	tr
Ukrainian	uk
Vietnamese	vi

Pre-downloading Model

To avoid delays during agent initialization, you can pre-download the Hugging Face model:

You can pre-download a specific language model:

from videosdk.plugins.turn_detector import pre_download_namo_turn_v1_model

# Download the English model before the agent runs
pre_download_namo_turn_v1_model(language="en")

Or pre-download the multilingual model:

from videosdk.plugins.turn_detector import pre_download_namo_turn_v1_model

# Download the multilingual model
pre_download_namo_turn_v1_model()

Additional Resources

The following resources provide more information about VideoSDK Turn Detector plugin for AI Agents SDK.

SDK Reference

GitHub Repository

Python Package