Model Details

• Model name: Audio-Interaction
• Task: Streaming audio-conditioned text generation (audio in, text out)
• Audio encoder: Qwen2.5-Omni audio tower (chunk-wise)
• Audio framing: 16 kHz, padded to 0.4-second (6400-sample) boundaries; 10 encoder-output frames per chunk
• Decoding states: LISTENING (emits KEEP_SILENCE / TEXT_BEGIN) and SPEAKING (emits text until TEXT_END)
• Default sampling: temperature 0.3, top-k 3
• Default max new tokens: 4096 per session
• License: Apache-2.0

Repository Contents

text
Copy code
Audio-Interaction/
├── model-00001-of-00004.safetensors      # LM weights, sharded (≈4 GB each)
├── model-00002-of-00004.safetensors
├── model-00003-of-00004.safetensors
├── model-00004-of-00004.safetensors
├── model.safetensors.index.json          # Shard index consumed by safetensors loader
├── config.json                           # Top-level model config
├── generation_config.json                # Generation defaults
├── model_config.yaml                     # GPT config consumed by Config.from_file
├── hyperparameters.yaml                  # Training-time hyperparameters (reference)
├── tokenizer.json                        # Tokenizer
├── tokenizer_config.json
├── MiniOmni3_ChunkwisedEncoder.pth       # Audio encoder weights (Qwen2.5-Omni audio tower)
└── qwen25OmniConfig/                     # Audio-encoder config (nested: thinker_config.audio_config)

Intended Use

Audio-Interaction is intended for streaming conversational agents that need to react to audio as it arrives — for example, voice assistants that may interject mid-utterance, alarms that respond to ambient sound, or low-latency dialogue systems where waiting for a full utterance before replying is too slow. The model is not a transcription system; it produces a conversational reply (or silence) rather than a verbatim transcript.

Quick Start

Installation

bash
Copy code
git clone https://github.com/xzf-thu/Audio-Interaction.git  # TODO: confirm repo URL
cd Audio-Interaction
conda create -n Audio-Interaction python=3.10 -y
conda activate Audio-Interaction
pip install -r requirements.txt

Download the checkpoint

From the Audio-Interaction project root, pull the weights into checkpoints/:

python
Copy code
from huggingface_hub import snapshot_download

snapshot_download(repo_id="zhifeixie/Audio-Interaction", local_dir="checkpoints")

snapshot_download is the recommended path — it pulls every file, resumes on interruption, and is the only way the download counter on this page advances. Please avoid git clone of the HF repo or the web "Download" button if you want your run reflected in the stats.

Python Usage

python
Copy code
from src.miniomni3.generate.run import run_inference

run_inference(
    checkpoint_dir="checkpoints",
    audio_paths=["/path/to/audio.wav"], # offline mode: one round per path
    device="cuda:0",                    # or "mps" / "cpu"
)

For interactive use, omit audio_paths and run_inference will prompt for an audio path each round:

python
Copy code
run_inference(checkpoint_dir="checkpoints", rounds=5, device="cuda:0")

Streaming Protocol

A single session looks like:

text
Copy code
[system prompt tokens]
  ┌─── LISTENING ───┐
  │ AUDIO_BEGIN PAD*10 ASSISTANT  →  KEEP_SILENCE          (keep listening)
  │ AUDIO_BEGIN PAD*10 ASSISTANT  →  TEXT_BEGIN EMOTION    (start replying)
  └─────────────────┘
  ┌─── SPEAKING ────┐
  │ … text tokens … TEXT_END                                (reply finished)
  └─────────────────┘
  ┌─── LISTENING ───┐  (next audio chunk)
  …

The model is trained to emit at most one TEXT_BEGIN per audio chunk. Each assistant turn begins with TEXT_BEGIN, followed by an emotion token, the reply tokens, and TEXT_END. Turns starting with KEEP_SILENCE indicate the model chose not to respond to that chunk.

Training Summary

Evaluation

Limitations

• The model produces text, not speech. Pair it with a TTS system for end-to-end voice interaction.
• Audio must be 16 kHz mono; non-conforming inputs are resampled by whisper.load_audio and padded to 0.4-second boundaries before encoding.
• Decisions are made at 0.4-second granularity (one encoder chunk), which sets a floor on response-onset latency.
• Trailing partial audio chunks shorter than 10 encoder frames are dropped before generation.

Citation

bibtex
Copy code
@misc{xie_miniomni3,
  title  = {Audio-Interaction: Streaming Audio-In, Text-Out Conversational Modeling},
  author = {Zhifei Xie and collaborators},
  year   = {2026},
  note   = {Preprint in preparation}
}

Acknowledgements

Audio-Interaction builds on the Qwen2.5-Omni audio encoder. We thank the Qwen team and the maintainers of OpenAI Whisper for the audio-loading utilities used in this project.