v0.1.34 音频增强与回放体验优化

本版整理自 v0.1.33 以来的音频增强、播放下载与处理选项改进。

• 音频增强统一为一个默认开启的总开关，高级设置可分别调整人声增益（0–200%，默认 100%）与 GTCRN 降噪混合（0–100%，默认 50%），并一键恢复服务端默认力度。力度仅影响新任务，0 跳过对应操作。
• 保留 GTCRN 50% 默认混合，撤下停顿门控并使用跨块连续增益，减少背景音随词句突然起落。降噪与增益分别记录状态、独立回退；不保证消除所有背景声，更高力度可能影响自然度或识别率。
• 音频／视频任务支持下载已保存的原音 WAV（16 kHz 单声道 PCM16；视频下载音轨），试听音轨与原音／增强音下载分组，切换保留播放位置与状态，手机分行显示。缺失的历史音轨不自动补生成；原上传文件仍通过 source 接口下载。
• “单任务加速”“音频增强”改为上下排列的可访问开关，直接展示已存力度、自定义／跳过及组件可用状态；TTS 文本／文档共用加速样式与偏好。更新中英文核心/API/升级文档、Swagger 示例及桌面／手机截图。

API 推荐 audio_enhancement=auto|off；兼容独立布尔字段，不能混用（422）。策略 6 保存两项力度、控制及后端，历史策略 ≤5 重放／重试不升级；向历史幂等键加入力度返回 409，需新 key。能力范围与默认值以 /api/v1/capabilities 为准。原音附件入口为 GET /api/v1/jobs/{job_id}/playback?download=true，支持 Range；默认回放行为不变。

从 v0.1.33 升级：数据迁移：无，无需额外升级备份，不改写历史请求、结果、媒体或声纹。ASR 浏览器偏好 v4 → v7 保留总开关并补默认力度；开发版 v5/v6 迁移见升级说明。存储期限、退出登录保留与草稿清理不变，TTS 偏好仍为 v2。更早版本须覆盖中间迁移；跨 v0.1.30 的 v13 → v14 仅需停服后的 SQLite 一致快照，无需复制媒体。完整验收后清理本次临时升级备份。

Qwen ASR/TTS 模型、隔离环境和依赖锁不变。setup asr/all 增加固定 revision/checksum 的实验性 UL-UNAS-DNS3；其公开 CER 验证未通过，GTCRN 仍是默认。按升级指南等待队列空闲、保留配置并停服更新，准备所需模型、重建前端，再刷新并验收 API、ASR/TTS supervisor 和 executor。运行时模型加载及音频处理均本地离线；本次发布未向外部服务上传敏感录音。本 Release 仅提供源码，不捆绑模型、运行环境或用户数据。

English: This release combines automatic denoising and quiet-speech gain behind one switch, with independent advanced strengths and reset-to-server-default support. It retains the validated GTCRN 50% default, removes pause gating, and uses continuous gain across processing blocks. Original playback WAV downloads join enhanced downloads, with separate listening/download groups and mobile rows. ASR and TTS processing switches now share an accessible layout. Higher strengths can affect naturalness and ASR accuracy; background conversations may remain.

From v0.1.33 there is no database/media migration or extra upgrade backup. Published ASR preferences v4 migrate to v7, preserving the master value and adding default strengths; storage/draft lifetimes and TTS v2 remain unchanged. Historical replay/retry semantics remain fixed. Earlier upgrades must cover intervening migrations; v13 → v14 requires a stopped SQLite snapshot without copying media. Qwen checkpoints and dependency locks remain pinned. Setup also offers the pinned experimental UL-UNAS checkpoint, which failed the public CER gate and is not the default. Rebuild the UI and refresh all previously running service processes after the final source/tag state. Runtime inference stays offline. Source archives contain no model weights, runtimes or user recordings.

API · 增强与公开样本验证 / Enhancement and public validation · 升级与本机部署 / Upgrade and local deployment · 工作台 / Workspaces

验证 / Validation:

• Exact commit: f5b7fc6ad1181c952f6f63199b7a5473db85cb2d.
• Clean-source local checks: 606 backend passed / 17 skipped; 258 browser passed / 5 optional Office skips; typecheck, production build, dependency locks, 183 documentation links/anchors, and temporary-tag version contracts. Real local GPU ASR/TTS smoke tasks passed, including original/enhanced WAV downloads and Range. Existing paired public CPU/GPU enhancement evidence remains applicable because the pipeline is unchanged; defaults were byte-identical on all eleven public controls.
• main Linux · main Windows
• tag Linux · tag Windows