DuplexJev-B-Whisper-small-Falcon-H1-3B

A small DuplexJev connector: the Whisper-small encoder (frozen) feeds a frozen Falcon-H1-3B, and every typed question is read as a single-token distribution over its options, with zero decode steps and many questions per forward pass. Trained with the same recipe as the Qwen3-32B connectors released with the paper.

Quick start

pip install "duplexjev[speech]>=0.2.2"

from duplexjev import Decider, Question

d = Decider.from_pretrained("adventists-ai/DuplexJev-B-Whisper-small-Falcon-H1-3B")

d.decide("utterance_en.wav", [
    Question("turn", "Has the user finished speaking?", ["finished", "not finished"], lang="en"),
    Question("intent", "What does the user want?", ["climate", "media", "navigation", "phone", "chit-chat"], lang="en"),
], lang="en")
d.decide("utterance_zh.wav", [                            # Chinese speech: Chinese question and options
    Question("turn", "用户说完了吗？", ["说完了", "还没说完"], lang="zh"),
    Question("intent", "用户想做什么？", ["空调", "媒体", "导航", "电话", "闲聊"], lang="zh"),
], lang="zh")

Falcon-H1 is a hybrid attention + Mamba model. Recurrent layers carry state along the sequence, so duplexjev (≥ 0.2.2) turns off prefix packing for it automatically (mode="batch": one row per question). For GPU speed install the Mamba kernels: pip install mamba-ssm causal-conv1d (they must be built for your torch/CUDA); without them transformers falls back to a slow reference path.

Results

Leaderboard scores (see the GitHub README): main language (mean of qa100, ZJU-ML, Easy-Turn) 46.5, paralinguistics (mean of gender, emotion) 39.1.

Paper protocol: the evaluation prompts of the paper (single option order, one H200). duplexjev defaults: the package's own prompts with lang matched to the clip. Spoken QA with a small LLM is bounded by the LLM itself (see the transcript-reading row); use these connectors for short decisions and speaker cues rather than open knowledge questions.

Latency of one decision event (10 questions, one 4.5 s clip, duplexjev 0.2.2, plain PyTorch): 169.7 s on 8 CPU threads (Xeon Platinum 8558, fp32, not quantized); about 104 ms on one H200 (bf16, measured while the GPU was shared with a training job). Falcon-H1 is not CPU-friendly: without Mamba CPU kernels transformers uses a slow reference path, and each question needs its own row.

Training

Frozen encoder and LLM; only the connector is trained: R1 and R2 transcript distillation on two disjoint 0.5 M-utterance packs of the Ultravox v0.6 mixture (32k steps each, global batch 16). Code and scripts: https://github.com/adventists-ai/duplexjev.

Limitations

• Evaluated on read or acted speech and small test sets (100–800 items); not on streaming input.
• Small LLMs answer knowledge questions poorly even from the transcript.
• Emotion labels come from acted corpora; do not use the outputs to make decisions about individuals.
• The connector only works with the encoder and LLM listed above.

License

Apache-2.0 for the connector weights. Some training corpora (e.g. WenetSpeech, CoVoST 2) are licensed for non-commercial use only; check them before commercial use. The frozen models keep their own licences: Falcon-H1-3B (TII Falcon License 2.0) and the encoder (Apache-2.0).

Citation

@inproceedings{jin2027duplexjev,
  title     = {Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen {LLM} Hear Beyond the Transcript},
  author    = {Jin, Jie and Ma, Ziyin and Yin, Min and Chen, Jinyu and Song, Haigang and Pang, Zhikun and Zhang, Xiaowen},
  booktitle = {Submitted to IEEE ICASSP},
  year      = {2027}
}

Built by Adventists.ai. Claude (Anthropic) assisted with code.