manifestro/kezek-v1

Kezek v1

A mono recording of exactly two people → one track per person, with the original length and timing.

Overlaps and short replies ("yeah", "mm-hm") stay on the track of the person who said them. The two tracks always sum exactly to the input.

Kezek is built for preparing full-duplex dialogue data from conversational audio that exists only as a mono mix.

Code (MIT) · Demo · Paper (arXiv, link at release)

Usage

pip install kezek
kezek-separate conversation.wav --out separated/

from kezek import Kezek
kz = Kezek.from_pretrained()            # this repository
tracks, activity, info = kz(audio_16k)  # float32 mono at 16 kHz -> tracks [2, T]

A CUDA GPU with bf16 is required.

Model

Kezek decides who speaks when and who is who once for the whole recording (stage 1). A conditioned separator then carries out that decision (stage 2).

File	Component	Parameters
fingerprint.pt	speaker fingerprint: raw-audio speaker embedding, trained to ignore phase	10.1 M
stage1_pairnet.pt	PairNet: activity of the two voices over the whole recording (20 ms), guided by the fingerprint	11.8 M
separator.pt	KezekTSE-L: Conv-TasNet-style separator conditioned on the activity and both voice profiles; outputs sum to the input	19.6 M

Total: 41 M parameters, all trained from our own runs, with no frozen third-party model.

Inference defaults: activity threshold 0.3; the separator runs over the whole recording in 8 s windows with a 6 s hop.

Evaluation

The comparison is against pyannote → DialogueSidon, as published in the DuplexChat code, run without changes on the same recordings. Whole recordings are scored; what a system does not return counts as lost.

Set	System	Short replies kept ↑	Leak (dB) ↓	No-swap ↑	Coverage ↑	h / GPU-h ↑
Open Yap test, as recorded (121 conv.)	pyannote → DialogueSidon	0.687	−10.1	0.793	0.735	51
	Kezek v1	0.830	−13.9	0.934	1.000	231
CallHome English, unseen telephone (128 calls)	pyannote → DialogueSidon	0.711 / 0.798	−9.7	0.797	0.816	46
	Kezek v1	0.796 / 0.907	−11.8	0.859	1.000	232

How each column is scored:

• Open Yap: scored against the reference channels.
• CallHome: scored by the utterance annotation. Replies are given over the other's speech / in pauses.
• Leak: energy on a track while only the other person speaks.
• No-swap: share of recordings without swapped channels.
• Speed: one RTX PRO 6000 GPU.

On Open Yap test, clean audio:

• overlap SI-SDRi +10.61 dB;
• short replies 0.878;
• no-swap 0.967.

Limitations

• The remaining gap is quiet replies under a louder voice. On Open Yap test, as recorded, 0.83 of short replies are kept.
• Exactly two speakers, offline; no enrollment; no speech restoration. Kezek splits the audio and does not clean it.
• Training and quantitative evaluation are in English.
• On telephone speech with similar voices, channels can swap for part of a call: 14% of CallHome calls have at least one swapped window.

Training data

Data	Use	Terms
Open Yap 1K, build open-yap-1k_2026-11-04_wav_0e7bcf1e_0d3abb77	conversations: activity, separation	Open Yap 1K Data Use Agreement v1. Commercial use and open weights allowed. No Open Yap audio or derived tracks are distributed here.
Libri-Light (LibriVox)	synthetic overlaps, fingerprint	public domain audio
VoxPopuli	fingerprint	CC0. We acknowledge the European Parliament for creating and sharing these materials.
MUSAN	augmentation (noise, music)	CC BY 4.0
RIRS_NOISES	augmentation (room impulse responses)	Apache-2.0

CallHome and DuplexChat audio were not used for training.

Responsible use

Kezek only reassigns audio that is already in a recording; it does not synthesise speech. Users must have the right to process their recordings. Do not use Kezek to identify speakers or to clone a person's voice.

Licence

Weights: Apache-2.0. Code: MIT (github.com/Manifestro/kezek).

Citation

@misc{karl2026kezek,
  title  = {Kezek: Splitting Mono Two-Person Conversations into Per-Speaker Tracks for Full-Duplex Dialogue Data},
  author = {Karl, Bagzhan and Karl, Alizhan},
  year   = {2026},
  note   = {arXiv identifier to be added at release}
}

Please also cite Open Yap 1K (The Agentic Data Company), build open-yap-1k_2026-11-04_wav_0e7bcf1e_0d3abb77.