Speaker Separation说话人分离
A conventional separation model may split one speaker across different tracks, merge two speakers into the same track, or severely degrade speech quality. Such low-quality separation is effectively unusable for training full-duplex speech models, where stable speaker identity and clean, time-aligned signals are fundamental requirements.
传统分离模型可能把同一个人拆到不同音轨、把两个人合入同一音轨，或者严重破坏语音质量。这类低质量分离结果基本无法用于全双工语音模型训练，因为稳定的说话人身份、清晰的人声和准确的时间对齐都是最基础的数据要求。
After analyzing and comparing the strengths of several models, we combine them into a coordinated workflow. VAD first partitions hour-scale source audio at natural boundaries into approximately 15-minute parts before diarization, improving Diarizen's speaker-identification performance. After music screening and low-SNR filtering, a greedy dynamic-window search captures high-quality reciprocal two-speaker segments that are as long as possible. This both removes long monologues from the training set and creates better input conditions for Sidon: alternating turns and sufficient speech from each person allow the separator to perform more reliably. After separation, speaker-consistency checks detect cases where one person is split across tracks or two people are merged into one track; affected chunks are automatically separated again to repair the result.
在分析并比较多个模型的特性后，我们将各自优势组合为一套协同工作的流程。首先利用 VAD 沿自然边界将小时级原始素材切分为约 15 分钟的片段，再交给 diarization 处理，从而提高 Diarizen 的说话人识别性能。完成音乐筛查和低 SNR 过滤后，系统通过动态窗口进行贪心搜索，捕获尽可能长、质量足够高且有来有回的双人对话片段。这一步既过滤了训练数据中的大段单人独白，也为 Sidon 分离创造了更理想的输入条件：两位说话人交替发言，并且各自拥有充分的语音时长，Sidon 才能发挥更稳定的分离性能。分离完成后，系统继续执行说话人一致性检查；一旦检测到同一个人被拆入不同音轨，或两个人被合入同一音轨，便会自动重新分离，对该片段进行修复。
SepFormer separates the selected chunk directly, without a downstream identity-consistency gate.SepFormer 直接分离选定片段，缺少后续的说话人身份一致性质量门。
OUR pipeline combines dynamic two-speaker windows, Sidon separation, identity-consistency checks, retry, and light background-noise suppression.OUR 管线结合动态双人窗口、Sidon 分离、身份一致性检测、自动重试和轻量底噪过滤。