Indic Multilingual Turn Detector

Indic Multilingual Turn Detector

Reads speech-to-text transcripts to judge whether a caller has finished their turn, in English, Hinglish and seven Indic languages. For voice agent and call-center developers.

Made by tarun7r on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This text model judges whether a person has finished their turn in a voice conversation. It reads the transcript from a speech-to-text system plus the conversation so far, then predicts if the speaker is done. That fills the gap left by voice activity detection, which notices silence but not meaning.

Who it suits: It suits developers of real-time voice agents and teams building agent assist for call centers, especially where callers speak Hindi, Hinglish or another Indic language.

Industry | Who uses it | How they use it | Example
Contact centers | Agent assist builders | Score live transcripts to tell when a caller is done | Hold the turn while a caller reads out a phone number
Voice AI | Voice agent developers | Use meaning, not a silence timer, to decide when to reply | Wait through a Hinglish sentence that ends on a dangling word
Voice platforms | LiveKit developers | Swap it in for LiveKit's text turn detector | Cover Indic languages beyond Hindi

Teams that work from raw audio only, or that need to identify who is speaking, can skip it.

Where to use it:

• Real-time voice agents that get transcripts from a speech-to-text pipeline
• Call-center agent assist tools
• LiveKit agents, through the EOUTurnDetector helper
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Decide when the agent may reply | Voice agent developers | A probability that the turn is complete
Hold through pauses and trailing words | Call-center tool builders | A signal to keep waiting while someone thinks
Replace LiveKit's older text detector | LiveKit developers | The same interface, with seven Indic languages covered

How to set it up:

• Python with pip, to download the files
• A speech-to-text pipeline that produces transcripts
• About 2 GB of disk for the ONNX graph
• For LiveKit, the EOUTurnDetector helper, which implements the MultilingualModel interface

1. Install the Hugging Face hub client.
2. Download the model files.
3. Load the ONNX graph, the one the thresholds were calibrated on.
4. Read the per-language thresholds from languages.json.

pip install -U huggingface_hub
hf download tarun7r/indic-turn-detector

How to use it: Feed the tokenized transcript to the ONNX graph as input_ids. It returns a prob value for every position, and you use the last one. Judge that value against the threshold stored for the language, choosing the safe or the balanced profile.

Tool | What it does
model.onnx | The fp32 graph used at runtime, taking input_ids and returning prob
languages.json | Holds safe and balanced thresholds for each language
model.safetensors | The PyTorch weights, with the eou_head tensors
EOUTurnDetector | A helper that plugs the model into LiveKit through MultilingualModel

Example: A caller starts reading out a phone number, and the transcript stops after the first few digits. The detector scores that text as an unfinished turn, so the agent keeps listening instead of cutting in. It holds the turn open until the number is complete.

For newcomers and experts: Many voice agents interrupt people because they only listen for silence. This model reads the words instead and judges whether the sentence sounds finished. A good first try is to download the files and score a short transcript that trails off mid-thought.

Under the hood it is Qwen2.5-0.5B-Instruct fine-tuned with an end-of-utterance head, with 494M parameters of which 179M were trainable. Its training data mixed MultiWOZ-style dialogues, TOP-style slots and generated negatives built from continuations and prefixes. Thresholds are calibrated on fp32 scores, and the ONNX graph matches those scores to within 1e-6. Verb-final languages are the hard case, because the predicate arrives last, right where silence-based cuts happen.

Speech-to-text produces a transcript
Tokens go in as input_ids
ONNX graph returns prob per position
Last position is checked against a threshold
Agent holds or takes its turn

Good to know:

• Languages: English, Hinglish and seven Indic languages, which are Hindi, Telugu, Tamil, Marathi, Bengali, Gujarati and Punjabi
• Size: 494M parameters, and about 2 GB for each of the ONNX and PyTorch files
• License: Apache 2.0
• Cost to run: none named, though you supply your own speech-to-text
• Data it reaches: the transcript text you pass in. No audio is processed, sent or kept, and each turn is scored on its own
• Not for: speaker identification, since it holds no voice biometrics or identity signal

On Hugging Face, 13 November 2026

• Task: text classification
• Library: transformers
• Parameters: 494M
• Weights: about 2 GB in F32; about 247 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 81
• Likes: 3
• License: apache-2.0
• Languages: en, hi, bn, gu, mr, pa, ta, te
• Based on: Qwen/Qwen2.5-0.5B-Instruct
• Tags: transformers, onnx, safetensors, qwen2, text-generation, voice-ai, voice-agents, turn-detection, end-of-utterance, end-of-turn, livekit, conversational-ai

Links

• Model: https://huggingface.co/tarun7r/indic-turn-detector
• Files: https://huggingface.co/tarun7r/indic-turn-detector/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download tarun7r/indic-turn-detector

Then follow the usage in its model card. Nothing in this kit runs by itself.