DuplexJev-4B-Para

DuplexJev-4B-Para

Takes speech and text in, answers in text, and judges when a speaker has finished talking. Built for voice-agent developers who want a small audio model that also reads emotion and stress.

Made by adventists-ai 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 is a 4.2B-parameter audio model that accepts speech and text and replies in text. Its main strength is live turn-taking, meaning it can tell when the person speaking has finished. It also handles emotion, stress, pauses, coughs, sighs and laughter, and speaks about eight languages including Chinese, English and Japanese.

Who it suits: It suits developers who build voice agents or speech tools and want a small model that listens, not only transcribes.

Industry | Who uses it | How they use it | Example
Voice agents | Voice assistant developers | Decide when a caller has stopped talking | Send audio and the last assistant message, then read the finished decision
Speech research | Audio researchers | Test emotion and stress recognition on speech clips | Ask which word in a sentence is stressed
Customer calls | Call tool builders | Detect pauses, laughter or coughs in a conversation | Flag laughter in a recorded Chinese chat
Product teams | Prototype builders | Try a small speech model before moving to the 32B one | Compare answers on the same audio

If you need commercial use or strong transcription in Japanese, Korean, Vietnamese, Thai or Arabic, you can skip it.

Where to use it: It runs as a vLLM server on your own machine, and voice-agent frameworks connect to it through plugins.

• vLLM, as a local server
• Pipecat voice agents, through the pipecat-duplexjev plugin
• LiveKit voice agents, through the livekit-plugins-duplexjev plugin
• A GPU machine with about 10 GB of memory

Use case | Who | What they get back
End-of-turn detection | Voice agent builders | A decision on whether the user has finished speaking
Emotion and stress checks | Speech researchers | Labels for emotion, stressed words and pauses
Sound detection | Audio tool developers | Answers about coughs, sighs and laughter in a clip

How to set it up: You need Python with pip and a GPU with room for about 10 GB. Version 1.4 needs the duplexjev-vllm plugin at 0.4 or newer, or duplexjev at 0.7 or newer.

1. Install vLLM with audio support and the DuplexJev plugin.
2. Start the server with the model name.
3. To get the older version, add the revision flag with the version name, for example v1.3.
4. Optionally, download the files to disk first with the Hugging Face tool.

pip install "vllm[audio]>=0.29" "duplexjev-vllm>=0.4"
vllm serve adventists-ai/DuplexJev-4B-Para --max-model-len 4096

How to use it: You send the user's audio as the user turn, using the audio placeholder, together with ordinary chat history, then sample as usual. The decision prompts match those of DuplexJev-4B-Turn. For stress and non-verbal questions, always offer a none option.

<|audio|>

Tool | What it does
vllm serve | Starts the model as a local server
pipecat-duplexjev | Adds live end-of-turn detection to a Pipecat voice agent
livekit-plugins-duplexjev | Adds live end-of-turn detection to a LiveKit voice agent

Example: A voice agent has just said something, the caller pauses, and the app sends the caller's audio plus the agent's last message. The model answers whether the caller has finished, so the agent knows if it should reply now or keep listening.

For newcomers and experts: Most speech tools turn audio into text and stop there. This one listens to the voice itself, so it can tell a finished sentence from a pause. A first small thing to try is serving the model and asking it about one short clip.

The design joins a Qwen3-ASR-0.6B encoder, a connector and Qwen3-4B. Version 1.4 hears the whole clip and was retrained in one pass, with a turn-taking stage added through a merged LoRA. The card reports end-of-turn accuracy rising from 71.0 to 92.9 percent on the Smart Turn v3.2 test set.

Caller audio and last message go in
The encoder turns audio into tokens
Qwen3-4B reads the audio and text
The model replies in text
Your app acts on the decision

Good to know: The card compares versions 1.3 and 1.4 on its own tests, and the newer one scores higher in each group it lists.

Title: End-of-turn accuracy on the Smart Turn v3.2 test set
Version 1.3: 71.0 percent
Version 1.4: 92.9 percent

• Limits: for Japanese, Korean, Vietnamese, Thai and Arabic transcription, the 32B model is much stronger.
• Self-description: ask it about itself and it may name skills it does not have, such as vision or a voice.
• Speaker judgments are coarse. Do not use them to identify or verify people, or to make decisions about individuals.
• License: CC BY-NC 4.0, so use stays limited to non-commercial work.
• Cost to run: none named. It needs about 10 GB of GPU memory.
• Data it reaches: the audio and text you send to your own server.
• Not for: commercial use, identifying people, or decisions about individuals.

On Hugging Face, 13 November 2026

• Task: audio text to text
• Library: vllm
• Parameters: 4.2B
• Weights: about 8.4 GB in BF16; about 2.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 363
• Likes: 6
• License: cc-by-nc-4.0
• Languages: zh, en, ja, ko, vi, th, ar, ru
• Based on: adventists-ai/DuplexJev-4B-Turn
• Tags: vllm, safetensors, ultravox, duplexjev, speech-to-decision, spoken-dialogue, full-duplex, turn-taking, end-of-turn, paralinguistics, nonverbal, speech-llm

Links

• Model: https://huggingface.co/adventists-ai/DuplexJev-4B-Para
• Files: https://huggingface.co/adventists-ai/DuplexJev-4B-Para/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download adventists-ai/DuplexJev-4B-Para

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