FastDialogueSidon

Weights for FastDialogueSidon, a fast one-step model for full-duplex two-speaker dialogue separation and joint restoration: up to 120 s of 16 kHz mono dialogue in, two separated 24 kHz speaker tracks out. It combines a one-step DiT head (a single DiT step per window instead of the 30 diffusion steps of DialogueSidon) with a lightweight decoder (a frame-rate SAN-WaveNeXt2 (WaveFitNeXt) latent decoder, 30 M parameters, in place of DialogueSidon's DAC decoder). Every module — encoder, DiT head and decoder — is trained for FastDialogueSidon. Inference code with a Triton backend (no engine build) and a TensorRT backend lives in higurashi-ai/FastDialogueSidon.

Model

file	params	what
encoder/model.safetensors	314.6 M	13-layer w2v-BERT 2.0-architecture encoder with relative_key attention, + 2 speaker latent heads
diffusion_head/model.safetensors	88.7 M	one-step DiT head (8 layers, hidden 768)
vocoder/model.safetensors	30.1 M	lightweight SAN-WaveNeXt2 (WaveFitNeXt) latent decoder (2 iterations, frame-rate ConvNeXt blocks)
stats.safetensors	—	per-channel latent normalization mean / std
config.json, */config.json	—	module hyper-parameters and the one-step update constants

All weights are fp32 safetensors (≈ 1.73 GB in total).

Usage

git clone https://github.com/higurashi-ai/FastDialogueSidon && cd FastDialogueSidon
hf download higurashi-ai/FastDialogueSidon --local-dir weights/
cd triton && pip install -e .
python scripts/run.py --bundle ../weights --wav dialogue.wav --output-dir out/   # any length <= 120 s

from fds_triton.runtime import TritonPipeline, load_waveform
pipe = TritonPipeline("weights/")
audio = pipe.separate(load_waveform("dialogue.wav"))   # [2, samples] @ 24 kHz

The TensorRT backend (trt/ in the code repository) builds its engines from the same weights: python scripts/build_engines.py --bundle ../weights --engine-dir artifacts/$(hostname)-trt10 --onnx-dir /tmp/fds-onnx.

Performance

Speed

Per-window wall clock on H100 (batch 1) against the original DialogueSidon — official release run as-is (fp32 PyTorch, DPM-Solver++ 30 steps + DAC) — on the benchmark inputs below:

	DialogueSidon	FastDialogueSidon (Triton fp16)	speed-up
OpenDialog, 30 s windows	0.53 s (RTF 0.0178)	6.2 ms (RTF 2.07e-4)	86x
CallFriend, 120 s windows	2.67 s (RTF 0.0222)	24.8 ms (RTF 2.07e-4)	107x

Benchmark

Protocol of the DialogueSidon paper (Sec. 4.1) with these deviations: CallFriend sessions cropped to 120 s (the paper uses 20 s), OpenDialog uses 100 English sessions of exactly 30 s, and intelligibility is scored with Qwen3-ASR-1.7B instead of Whisper large-v3. Both systems run on the same inputs. Neither CallFriend nor OpenDialog is used for training; both serve only as test sets.

OpenDialog (English, 100 × 30 s)

	WER (%) ↓	DNSMOS ↑	NISQA ↑
DialogueSidon	13.21	3.327	3.757
FastDialogueSidon (Triton fp16)	10.80	3.303	3.491

CallFriend (5 languages × 20 × 120 s)

p-CER (%) ↓	de	es	fr	ja	zh	all
DialogueSidon	11.56	15.39	21.02	16.36	20.93	16.54
FastDialogueSidon (Triton fp16)	11.47	16.89	21.25	14.49	19.46	16.52

	DNSMOS ↑	NISQA ↑	speaker sim. ↑	VAD acc. ↑
DialogueSidon	3.293	3.626	0.941	0.931
FastDialogueSidon (Triton fp16)	3.294	3.395	0.943	0.929

See the code repository for per-length latencies of both backends (Triton and TensorRT), where the speed-up comes from, and the numerical-accuracy comparison.

License

Apache License 2.0.

Citation

FastDialogueSidon technical report (Rikuto Kotoge): coming soon.

@inproceedings{nakata2026dialoguesidon,
  title={DialogueSidon: Recovering Full-Duplex Dialogue Tracks from In-the-Wild Dialogue Audio},
  author={Nakata, Wataru and Saito, Yuki and Yamauchi, Kazuki and Tsunoo, Emiru and Saruwatari, Hiroshi},
  booktitle={Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue},
  pages={1--12},
  year={2026}
}