v0.1.9-speaker

feat: temporal overlap mapping for multi-speaker evaluation

- Replace position-based remapping with temporal overlap-based remapping
- Add groundTruthTimeline with timestamps for 7-speaker meeting audio
- Add remapSpeakersByTimeline function using temporal overlap
- Remove unused remapSpeakersToCanonical function
- Auto-tune with new evaluation: 87.2/100, 49% similarity, 7 speakers

Co-Authored-By: AtomCode (deepseek-v4-flash) <noreply@atomgit.com>

v0.1.8-speaker — Multi-speaker auto-detect

v0.1.8-speaker

Breaking Changes

• Removed numSpeakers=2 constraint — FluidAudio now auto-detects speaker count via VBx clustering

Improvements

• Multi-speaker support: Works with 2–7+ speakers (tested with 7-speaker meeting)
• Auto-tuned parameters: clustering.threshold=0.7, segmentation.stepRatio=0.3, embedding.minSegmentDuration=1.0s
• Auto-tune function: Supports multi-speaker evaluation with Damerau-Levenshtein sequence matching and speaker remapping
• Evaluation: Updated scoring function handles any number of speakers (not just 2)

Test Results

• 7-speaker meeting (31 min): 7 speakers detected ✓, score 85.7/100

v0.1.7-speaker — Auto-tuned diarization

v0.1.7-speaker — AutoResearch 自动调优版

主要变更

• AutoResearch 式参数自动调优，33 次试验找到最佳参数组合
• 参数优化：threshold=0.6, stepRatio=0.15, minSegmentDuration=0.3
• 新增 --tune 标志，支持对新音频自动调优
• 说话人识别准确率大幅提升

调优关键发现

参数	旧值	新值
clustering.threshold	0.5	0.6
segmentation.stepRatio	0.2	0.15
embedding.minSegmentDuration	0.5	0.3
segmentation.minDurationOn	0.2	0.0
segmentation.minDurationOff	0.3	0.0
postProcessing.minGapDuration	0.05	0.1

下载

• ohr-speaker_macOS_arm64.zip — macOS Apple Silicon (M1/M2/M3/M4)

v0.1.6-speaker

ohr-speaker v0.1.6-speaker

ohr 增强版，集成 FluidAudio 离线说话人分割（Speaker Diarization）。

新增功能

• 🗣️ 说话人识别 — --speakers 标志启用声纹分割
• 📝 支持说话人标签的纯文本、JSON、SRT、VTT 输出
• 🖥️ 服务器模式支持 diarize 参数

安装

chmod +x ohr-speaker
./ohr-speaker --speakers 音频文件.wav

说明

• 100% 本地运行，无需联网
• 首次运行自动下载 FluidAudio 模型（~700 MB）
• 基于 ohr v0.1.6