From 7f1b9f976179aa3dc6aec04102ca303be5cc860e Mon Sep 17 00:00:00 2001 From: Hawk Lee Date: Mon, 23 Feb 2026 23:55:59 +0800 Subject: [PATCH] fix: FA stop consonant cutoff and tail allowance; bump version to 1.15.5 --- aiia_podcast_stitcher.py | 50 +++++++++++++++++++++++----------------- pyproject.toml | 2 +- 2 files changed, 30 insertions(+), 22 deletions(-) diff --git a/aiia_podcast_stitcher.py b/aiia_podcast_stitcher.py index b87b897..2035104 100755 --- a/aiia_podcast_stitcher.py +++ b/aiia_podcast_stitcher.py @@ -811,29 +811,37 @@ class AIIA_Podcast_Stitcher: if fa_entry: fa_start, fa_end = fa_entry['start'], fa_entry['end'] - # FA 精确起点 + 能量检测自然收尾（受 VAD 约束） + # FA 精确起点 cut_start = fa_start - cut_end = self._refine_cut_point(wav, sr, fa_end, - search_radius=0.15, direction="after") - # 用 VAD 约束尾部：cut_end 不应超过 VAD 检测到的实际语音结束 + + # 针对清辅音（如 PPT 的 T）或者呼吸尾音经常被 FA 强制截断的问题： + # 直接给予充分的自然发音衰减物理时间 (0.35s) + TAIL_ALLOWANCE = 0.35 + + if fa_results and sent_local_idx + 1 < len(fa_results): + next_fa = fa_results[sent_local_idx + 1] + if next_fa and next_fa['start'] > fa_end: + gap = next_fa['start'] - fa_end + # 延伸进入 gap，最多 TAIL_ALLOWANCE，且不超过 gap 的 60% (避免吃进下一句的气口) + extend = min(TAIL_ALLOWANCE, gap * 0.6) + cut_end = fa_end + extend + else: + # 如果下一句贴得很紧，保底向后略微缓冲 + cut_end = fa_end + TAIL_ALLOWANCE / 2 + else: + cut_end = fa_end + TAIL_ALLOWANCE + + # 用 VAD 验证尾部：寻找能量谷和 VAD 终点附近的合理落点 if use_vad: vad_ts = vad_timestamps_A if speaker == "A" else vad_timestamps_B if vad_ts: - _, vad_offset = self._refine_with_vad(fa_start, fa_end, vad_ts, search_margin=0.3) - # 取能量谷和 VAD 终点中更早的（避免切入静音区） - cut_end = min(cut_end, vad_offset) - - # 防止尾部扩展吃进下一句的 FA 起点 - if fa_results and sent_local_idx + 1 < len(fa_results): - next_fa = fa_results[sent_local_idx + 1] - if next_fa and cut_end > next_fa['start']: - # 取中点，但保底不低于 FA 原始终点（防止 FA 时间戳微重叠时过度截断） - cut_end = max(fa_end, (fa_end + next_fa['start']) / 2) - else: - # 最后一句：硬性上限 = FA 终点 + 0.3s - # 防止 CosyVoice/TTS 在音频末尾重复最后一句被误收 - MAX_TAIL_AFTER_FA = 0.3 - cut_end = min(cut_end, fa_end + MAX_TAIL_AFTER_FA) + # 放大 search_margin 容忍拖音 + _, vad_offset = self._refine_with_vad(fa_start, fa_end, vad_ts, search_margin=0.5) + # 如果 VAD 侦测到更长的尾音，则采纳更长的；但绝不早于我们的物理推断 + cut_end = max(cut_end, vad_offset) + + # 取能量谷微调：在估算好的 cut_end 附近找一个真正安静的帧切断，避免切在底噪波峰 + cut_end = self._refine_cut_point(wav, sr, cut_end, search_radius=0.10, direction="both") # 交叉验证：同时计算 VAD 和 Energy 的结果做对比 if use_vad and vad_timestamps_A is not None: @@ -874,9 +882,9 @@ class AIIA_Podcast_Stitcher: cut_start = self._refine_cut_point(wav, sr, cut_start, search_radius=0.15, direction="before") cut_end = self._refine_cut_point(wav, sr, cut_end, search_radius=0.10, direction="both") - # 应用 padding：cut_start 全额保留起音余量，cut_end 少量保留尾音衰减 + # 应用 padding：cut_start 全额保留起音余量，cut_end 增加较多尾音衰减 cut_start = max(0, cut_start - padding) - cut_end = min(len(wav) / sr, cut_end + padding * 0.3) + cut_end = min(len(wav) / sr, cut_end + padding * 0.8) # 防重叠：确保 cut_start 不早于同一说话人上一个片段的 cut_end if cut_start < prev_cut_end[speaker]: diff --git a/pyproject.toml b/pyproject.toml index 9d919ba..d14fee1 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,7 +1,7 @@ [project] name = "aiia" description = "Advanced AI Audio/Video toolkit for ComfyUI. Features Multi-Role Podcast/Dialogue Generation, High-Fidelity Voice Cloning (CosyVoice/VibeVoice/IndexTTS-2), TTS, Media Management, and efficient Video tools." -version = "1.15.4" +version = "1.15.5" license = {file = "LICENSE"} readme = "README.md" authors = [