Wave Subs 1.0.9

Wave Subs 1.0.9

Windows

• 用户名里有中文（或日文、韩文、俄文、带重音的字母）时，语音识别一开始就崩溃，退出码 3221226505。之前所有 Windows 版本都有这个问题，现在修好了（#1）
• VAD 语音检测同样修好。以后它万一跑不起来，质检里会明确提示，不再悄悄降低时间轴的精度
• 转换时不再弹出黑色的命令行窗口
• 画面文字识别在中文、日文、韩文 Windows 上可能完全用不了，开关也不显示，现在修好了

其它

• 识别没有产出结果时，报错里带上识别程序自己的日志，不再只说「文件不存在」

---

Windows

• Speech recognition crashed right at the start (exit code 3221226505) when the Windows user name contained Chinese, Japanese, Korean, Cyrillic or accented characters. Every earlier Windows version was affected; this is fixed (#1)
• Speech detection is fixed the same way. Should it ever fail to run, the quality report now says so instead of quietly lowering timing precision
• No more black console windows during a conversion
• On-screen text could be unusable on Chinese, Japanese and Korean Windows, with its switch hidden; this is fixed

Other

• When recognition produces no result, the error now includes the recognizer's own log instead of only "file not found"

Wave Subs 1.0.8

Wave Subs 1.0.8

设置

• 「自动检查更新」改叫「启动时检查更新」，小字说明「应用启动时进行一次更新检查」：它只在启动时查一次，之前的文案让人以为运行期间也会查

---

Settings

• "Check for updates automatically" is now "Check for updates at launch", with the note "Checks for a new version once when the app starts": it only checks once, at launch, and the old wording suggested it kept checking while running

Wave Subs 1.0.7

Wave Subs 1.0.7

翻译画面中的文字

• 转换设置里新增开关（需开启翻译）。用系统自带的文字识别（macOS 的 Vision、Windows 的 Windows.Media.Ocr）逐秒读画面，把招牌、便签、短信与聊天气泡、告示、文件、标题卡、人物名牌翻译出来，按原文的位置写进 ASS（SRT 放顶部）。不下载任何模型，24 分钟一集多花两三分钟，与语音识别并行
• 译文贴着原文放，尽量不遮挡：贴原文正下方 → 放到原文左右的空处 → 挪到画面顶部 → 盖在原文上，四种排法按顺序试。让位分三轮，两轮都没位置就把字号压小再来一遍——宁可字小，也别盖住原文。只有整屏的邮件、短信、文件才铺一块深色底板把原文换掉
• 底部永远属于对白：同一时刻有对白字幕时，画面文字绝不进入字幕占用的高度，按对白的实际行数让位
• 看得见的每一块都有译文：同屏放得下八条，和对白说的是同一句也照出，实在排不下就压到最小字号也要排出来
• 不该出的不出：片头片尾名单（连片尾「角色名／声優名」那一段一并认出）、片源自带的烧录字幕、电视台台标与频道水印、地图番号与书架书脊这类密集小字、走廊里反复入镜的门牌、被 OCR 读坏的颜文字、以及译文和原文一样的（英文照抄、中日同形、只做了简繁转写的日剧人名字幕）
• 编辑器：语音字幕与画面文字分两个页签，各自增删改；预览里画面文字按排好的位置与字号叠在画面上

翻译

• 术语表现在也管画面文字。之前只喂给对白翻译，同一个人名在两条轨道上会对不上；术语表改了，对白与画面文字一起重译
• 小模型把长句只译成开头几个字：用 1.7B 翻日语纪录片时整集译文都只有原文头一个短语（「今年6月」「受邀的是」），根因是提示词里要求先照抄原文开头几个字，小模型把译文也只译那几个字。提示词已改
• 小模型翻出来的整批译文不再丢：1.7B 常把结果输出成一行一个数组、或者中途截断，之前解析不了就整批 20 条作废，退到逐条重试——而逐条时小模型最爱把原文照抄回来，抄回来的又没被拦住直接进了字幕（双语模式同一句日文出两遍）。现在能救的都救回来，每一轮都拦照抄。一集 NHK 纪录片（Tiny + 1.7B）里没译文或照抄的对白从 98 条降到 56 条
• 小模型不再把原文照抄成译文：1.7B 翻日语时第一轮就有两成左右的条把日文原样抄进译文，听写有错字的句子更是整批照抄。提示词现在按源语言、目标语言各带一句示例，外加「即使原文是语音听写、有错字或不通顺，也按最可能的意思译出来」；最难的几批实测第一轮真正译出来的比例从 40% 到 88%
• 同一个专名不再有两种写法。翻译是分批做的，批与批之间没有记忆，一部片里同一个名字会出现「韦伯／威伯」两种写法。收尾按源文里反复出现的专名把译文分组，把少数写法改成多数写法
• 同一个文件跑两遍出同一份字幕：画面文字的判别改成贪心解码加固定随机种子

其它

• 模型页按语言给参考：每个识别模型标出英语、欧洲语言、日语·韩语·中文三组的意思保留率和档位（推荐 / 可用 / 勉强 / 不建议），列表上方一句「按语言挑模型」。数字来自三部片各 30 分钟的实测盲评
• 检查更新：设置页「关于」里多一行，启动后自动查一次，也能手动点；有新版本就给下载按钮和更新说明，可以跳过某个版本，也可以关掉自动检查。用 Homebrew 或 Scoop 装的会给对应的升级命令
• 自动判断语种更稳：不再只看开头 30 秒（片头音乐会被判成英语，整集识别成英文胡话），改为在人声最密的几处各取 20 秒投票
• 字幕样式统一：画面文字原来是粗体、对白是常规体，同一屏上看着像两种字体，现在统一

Windows

• 语音识别不再随包带 OpenBLAS：之前随包的官方 whisper-blas-bin 预编译包里的 OpenBLAS 在一部分机器上一加载模型就崩（「whisper-cli 识别失败（退出码 3221225477）」）。换成不带 BLAS 的官方包，常用模型速度不变，装到磁盘上少 51 MB
• 模型页按 Windows 机器判断：CPU 型号按实际读，不再一律写成「这台 Mac：Intel」；适配标签与要求文案不再照搬 Mac 的规则
• 任务失败时可以一键复制完整日志
• 深色模式下的下拉菜单能看清了：之前语种、模型这些列表在 Windows 上整张看不见，只有选中的那一行能读

首次使用

• 刚装好、一个模型都没有时，转换页直接给出这台机器的推荐模型和「下载并继续」，进度在原地走，下完直接开始；拖了文件再发现没模型也不用离开这一页。本地翻译缺模型同样处理
• 转换设置拆成「识别」「翻译」两个框：识别框是字幕来源 → 识别模型 → 语言，翻译框是翻译服务（不翻译 / 本地 / 云端）→ 模型 → 翻译成 → 画面文字 → 字幕内容，服务没就绪之前后面几行不出；字幕格式放在「开始」旁边。批量页同样
• 没有模型时不再摆一排没意义的选项：识别框里只留「字幕来源」和下载提示，翻译框里只留「翻译成」「翻译服务」和下载提示，模型下好了选项再出来
• 拖进来的文件在切换标签时不再丢，只有点「取消」才回到空态
• 按钮各态重画：主按钮在浅色模式下 hover 时不再变成一块看不清字的白板；键盘焦点有了统一的环

已知限制

• Windows 上的画面文字用系统自带的 OCR，要先在设置里装好片中语言的语言包（含「光学字符识别」）；识别质量不如 macOS，竖排日文基本读不出
• 名单之外的一串人名仍会被译出来：话剧海报上成排的演员名，以及和制作名单隔了半分钟先单独打出来的主演名。
  前者和纪录片里逐个出现的流程图标签在结构上分不开；后者靠放宽名单的邻接阈值能捞回来，代价是片头社团招牌、片尾正文会被误吃，不划算
• 本地 8B 模型对专有名词不稳（机构名、历史术语会译错或前后不一致）。术语表是可靠的办法
• 片源自己把画面文字的译文烧在画面上时（部分字幕组的做法），会和本程序的译文并存

---

Translate on-screen text

• A new switch in conversion settings (requires translation to be on). The text recognition built into the OS (Vision on macOS, Windows.Media.Ocr on Windows) reads the picture once a second, and signs, notes, text messages and chat bubbles, notices, documents, title cards and lower-third name plates are translated and written into the ASS at the position of the original (SRT places them at the top). No extra model to download; a 24-minute episode takes two to three minutes longer, running alongside speech recognition
• The translation is placed next to the original, covering it only as a last resort: directly below → in the free space to its left or right → at the top of the frame → over the original, tried in that order. Yielding runs in three rounds, and when two rounds find nothing the font is reduced and the search repeats — smaller type is better than hiding the original. Only a full screen of text (an e-mail, a message thread, a document) is replaced with an opaque plate
• The bottom belongs to the dialogue: when a dialogue subtitle is on screen, on-screen text never enters the height it occupies, measured from its actual line count
• Everything visible gets a translation: up to eight at once, kept even when the dialogue says the same thing, and placed at the minimum font size rather than dropped when nothing else fits
• What stays out: opening and closing credits (including the character/voice-actor list at the end), subtitles already burned into the video, TV station logos and watermarks, dense small labels such as map unit markers and book spines, door signs that keep recurring through a film, emoticons garbled by recognition, and translations identical to the original
• Editor: speech subtitles and on-screen text are separate tabs, each editable; the preview overlays on-screen text at its computed position and size

Translation

• The glossary now applies to on-screen text as well. It previously reached only the dialogue, so the same personal name could differ between the two tracks; changing the glossary now retranslates both
• Small models translated only the first few words of a long line: with the 1.7B model a Japanese documentary came out as one short phrase per line ("This June", "Those invited were"). The cause was the alignment anchor in the prompt — copy the first four characters of the source, then translate — which small models applied to the translation too. The prompt now says so explicitly
• Batches from small models are no longer thrown away: the 1.7B model often returns one array per line or gets cut off mid-output, and the parser accepted only a single complete array, so a whole batch of 20 good translations was discarded and retried line by line — where small models most often copy the source back, and those copies slipped into the subtitles unchecked (the same Japanese line twice in bilingual mode). Every salvageable row is now kept, and copied-back source is rejected in every round. In one NHK documentary episode (Tiny + 1.7B), dialogue lines with no translation or a copied-back source fell from 98 to 56
• Small models no longer hand the source back as the translation: with the 1.7B model about a fifth of Japanese lines came back copied verbatim on the first pass, and lines with transcription errors were copied as whole batches. The prompt now carries a two-line example in the source and target languages plus an instruction to translate the most plausible meaning even when the transcript is garbled; on the hardest batches the first-pass yield of real translations went from 40% to 88%
• One proper noun, one spelling. Translation runs in batches with no memory between them, so a single name could come out two ways in one film. A final pass groups translations by the recurring proper noun in their source and rewrites the minority spelling
• The same file twice gives the same subtitles: on-screen text judging now uses greedy decoding with a fixed seed

Other

• The Models page now gives per-language guidance: each recognition model shows meaning retention for English, European languages and Japanese · Korean · Chinese with a verdict (recommended / usable / marginal / not advised), plus a one-line "choosing by language" note above the list. The figures come from blind-judged 30-minute samples of three films
• Check for updates: a new row under Settings › About checks once after launch and on demand; when a newer version exists it offers a download button and the release notes, with the option to skip a version or turn automatic checks off. Copies installed with Homebrew or Scoop get the matching upgrade command instead
• Steadier language detection: instead of only the first 30 seconds (where opening music could be read as English and turn a whole episode into English gibberish), five 20-second windows are sampled where speech is densest and vote
• Consistent subtitle styling: on-screen text was bold while dialogue was regular, which read as two different typefaces on the same frame

Windows

• Speech recognition no longer ships with OpenBLAS: the bundled official whisper-blas-bin build crashed on some machines as soon as the model loaded ("whisper-cli failed (exit code 3221225477)"). Switched to the official non-BLAS build — same speed on the common models, and 51 MB less on disk
• The Models page now judges the actual Windows machine: the CPU model is read as-is instead of always showing "This Mac: Intel", and the fitness tags and requirement lines no longer apply the Mac rules
• A failed task now has a one-click "Copy log" button
• Dropdown menus are readable in dark mode: the language and model lists used to render as an unreadable pale list on Windows, with only the selected row legible

First run

• With no model installed yet, the Convert page recommends a model for this machine and offers "Download and continue" right there — progress shows in place and you can start as soon as it finishes; dropping a file first no longer sends you off to the Models page. The same applies when local translation has no model
• Conversion settings are now two boxes, Recognition and Translation: source → recognition model → language, and service (none / local / cloud) → model → translate-into → on-screen text → subtitle content, the later rows appearing only once the service is ready; the subtitle format sits next to Start. Same on the Batch page
• Options that need a model stay hidden until one is installed: the Recognition box shows only the source and the download prompt, the Translation box only the target language, the service and its prompt; the rest appears once the model is there
• A dropped file survives switching tabs; only Cancel returns to the empty state
• Button states redrawn: the primary button no longer turns into an unreadable pale block on hover in light mode, and keyboard focus has a consistent ring

Known limitations

• On Windows, on-screen text uses the built-in OCR: install the language pack for the spoken language first (with "Optical character recognition"); recognition is weaker than on macOS, and vertical Japanese is mostly missed
• A run of personal names outside the credit roll still gets translated: rows of actor names on a theatre poster, and lead billing shown half a minute before the staff roll.
  The first is structurally indistinguishable from the diagram labels a documentary reveals one at a time; the second could be caught by wi...

Wave Subs 1.0.6

Wave Subs 1.0.6

模型下载（中国大陆根治）

• hf-mirror.com 已不再代理文件，只是跳回 huggingface.co，所以之前的"镜像"在大陆实际上下不了。现在全部模型都能从 ModelScope（魔搭，阿里云 CDN）下载：whisper 六个模型用两个与官方逐字节一致的镜像仓，Qwen 用官方仓；系统语言为简体中文或时区在中国时 ModelScope 排第一位，其它地区仍先走 huggingface.co
• 下载完成后核对 sha256，不一致就删掉重来，不会把来路不明的文件当模型用
• Silero VAD 模型随安装包自带，时间校正不再依赖联网下载
• 下载失败提示改为"所有下载来源都连不上"，并列出可复制的各来源地址，可用浏览器或下载工具自行下载后放进模型文件夹

Model downloads (fix for mainland China)

• hf-mirror.com no longer proxies files (it just redirects back to huggingface.co), so the "mirror" added in 1.0.3 never worked there. Every model can now be downloaded from ModelScope (Alibaba Cloud CDN): the six whisper models from two mirror repos verified byte-for-byte against the official files, Qwen from its official repo. ModelScope is tried first when the system language is Simplified Chinese or the time zone is in China; elsewhere huggingface.co stays first
• Downloads are verified against their sha256 after completion; a mismatch is deleted and retried, so a tampered file is never used as a model
• The Silero VAD model now ships inside the app, so timing refinement no longer depends on a download
• A failed download says "none of the download sources could be reached" and lists every source URL with copy buttons; fetch the file with a browser or download manager and drop it into the model folder

---

下载 / Downloads

	文件 / File	说明 / Notes
macOS	Wave.Subs-1.0.6-arm64.dmg	Apple Silicon，已公证 / notarized
macOS	Wave.Subs-1.0.6-arm64-mac.zip	同上，ZIP / same, as ZIP
Windows	Wave.Subs.Setup.1.0.6.exe	x64 安装程序（未签名，SmartScreen 提示时点「更多信息 → 仍要运行」）/ x64 installer, not yet code-signed
Windows	Wave.Subs-1.0.6-win.zip	便携版 / portable

校验值见 SHA256SUMS.txt。包管理器 / package managers: brew install --cask jason-jm/wavesubs/wavesubs · scoop bucket add wavesubs https://github.com/jason-jm/scoop-wavesubs && scoop install wavesubs

Wave Subs 1.0.5

Wave Subs 1.0.5

转换过程

• 可以取消了：单文件的进度卡片和批量队列里正在跑的文件都有「取消」按钮，立刻停掉正在跑的抽音频 / 识别 / 翻译；批量里被取消的文件退回「等待中」，队列继续处理后面的
• 显示预计剩余时间：按当前阶段（识别、翻译）的推进速度估算
• 完成后显示这次识别用的设备，如「Apple M1」「NVIDIA GeForce RTX 3060 (Vulkan)」或「CPU」，一眼看出有没有用上 GPU

Windows

• 翻译走 GPU：随包的 llama-server 换成 Vulkan 构建，NVIDIA / AMD / Intel 显卡装有 Vulkan 驱动就用它，没有就自动回落 CPU；语音识别暂时仍在 CPU 上

模型下载

• 下载失败时给出具体原因，并列出该模型的全部下载地址供复制：用浏览器或下载工具下好、放进模型文件夹即可使用
• Qwen 翻译模型新增 ModelScope（魔搭）来源，官方源与镜像都连不上时自动尝试

Conversion

• Jobs can be cancelled: the progress card and the running item in the batch queue have a Cancel button that stops the audio extraction / recognition / translation immediately; a cancelled batch item goes back to "waiting" and the queue moves on
• Estimated time remaining, extrapolated from the pace of the current stage (recognition, translation)
• The result card shows which device did the recognition, e.g. "Apple M1", "NVIDIA GeForce RTX 3060 (Vulkan)" or "CPU", so you can tell at a glance whether the GPU was used

Windows

• Translation on the GPU: the bundled llama-server is now the Vulkan build. NVIDIA / AMD / Intel GPUs with a Vulkan driver are used, otherwise it falls back to the CPU automatically; speech recognition still runs on the CPU for now

Model downloads

• A failed download now reports the actual cause and lists every download URL for that model with copy buttons: fetch the file with a browser or download manager and drop it into the model folder
• Qwen translation models gain a ModelScope source, tried automatically when neither the official host nor the mirror is reachable

---

下载 / Downloads

	文件 / File	说明 / Notes
macOS	Wave.Subs-1.0.5-arm64.dmg	Apple Silicon，已公证 / notarized
macOS	Wave.Subs-1.0.5-arm64-mac.zip	同上，ZIP / same, as ZIP
Windows	Wave.Subs.Setup.1.0.5.exe	x64 安装程序（未签名，SmartScreen 提示时点「更多信息 → 仍要运行」）/ x64 installer, not yet code-signed
Windows	Wave.Subs-1.0.5-win.zip	便携版 / portable

校验值见 SHA256SUMS.txt。包管理器 / package managers: brew install --cask jason-jm/wavesubs/wavesubs · scoop bucket add wavesubs https://github.com/jason-jm/scoop-wavesubs && scoop install wavesubs

Wave Subs 1.0.4

Wave Subs 1.0.4

下载体验

• 点击下载后先显示「正在连接服务器…」，不再对着 0% 干等
• 官方源 3 秒没回应就并行连镜像，谁先回应用谁；赢过的来源记住并落盘，之后的下载和重启后都直接走它

Windows

• 界面语言优先读取系统「显示语言」设置，修复中文系统显示英文的问题；「设置 → 语言」下方列出系统报告的语言标签，方便对照
• 统一两个平台的窗口结构：顶部一条窗口栏，最小化 / 最大化 / 关闭画在栏里，内容从栏下面开始，不再压在卡片边框上

Downloads

• The button now shows "Connecting to server…" first instead of sitting at 0%
• If the official host does not answer within 3 seconds the mirror is tried in parallel and whichever responds first wins; the winning host is remembered on disk for later downloads and launches

Windows

• Interface language now follows the Windows "display language" setting first (fixes English UI on Chinese systems); Settings → Language lists the language tags the system reports
• One window layout for both platforms: a top title-bar strip hosts the minimize / maximize / close buttons, and content starts below it instead of colliding with the card border

---

下载 / Downloads

	文件 / File	说明 / Notes
macOS	Wave.Subs-1.0.4-arm64.dmg	Apple Silicon，已公证 / notarized
macOS	Wave.Subs-1.0.4-arm64-mac.zip	同上，ZIP / same, as ZIP
Windows	Wave.Subs.Setup.1.0.4.exe	x64 安装程序（未签名，SmartScreen 提示时点「更多信息 → 仍要运行」）/ x64 installer, not yet code-signed
Windows	Wave.Subs-1.0.4-win.zip	便携版 / portable

校验值见 SHA256SUMS.txt。包管理器 / package managers: brew install --cask jason-jm/wavesubs/wavesubs · scoop bucket add wavesubs https://github.com/jason-jm/scoop-wavesubs && scoop install wavesubs

Wave Subs 1.0.3

Wave Subs 1.0.3

模型下载在中国大陆可用了

• huggingface.co 连不上时自动切换到镜像 hf-mirror.com，成功一次后后续文件直接走镜像
• 下载改走系统网络栈，认系统代理设置；两边都失败时给出明确提示，不再是 "TypeError: fetch failed"

其它

• 界面语言检测更稳：修复部分 Windows 上中文系统显示英文的问题
• 官网：Windows 访客现在高亮并优先显示 Windows 版下载；下载区注明 SmartScreen 提示的处理方法

Model downloads now work from mainland China

• Falls back to the hf-mirror.com mirror automatically when huggingface.co is unreachable, and keeps using the mirror once it succeeded
• Downloads go through the system network stack and honour system proxy settings; a clear message replaces "TypeError: fetch failed" when neither host is reachable

Also

• More robust interface-language detection (fixes some Windows setups showing English on a Chinese system)
• Website: Windows visitors now get the Windows download highlighted first; the SmartScreen "More info → Run anyway" step is explained under the download

---

下载 / Downloads

	文件 / File	说明 / Notes
macOS	Wave.Subs-1.0.3-arm64.dmg	Apple Silicon，已公证 / notarized
macOS	Wave.Subs-1.0.3-arm64-mac.zip	同上，ZIP / same, as ZIP
Windows	Wave.Subs.Setup.1.0.3.exe	x64 安装程序（未签名，SmartScreen 提示时点「更多信息 → 仍要运行」）/ x64 installer, not yet code-signed
Windows	Wave.Subs-1.0.3-win.zip	便携版 / portable

校验值见 SHA256SUMS.txt。包管理器 / package managers: brew install --cask jason-jm/wavesubs/wavesubs · scoop bucket add wavesubs https://github.com/jason-jm/scoop-wavesubs && scoop install wavesubs

Wave Subs 1.0.2

Wave Subs 1.0.2

反馈更方便了

• 「帮助」菜单和「设置 → 关于」里可以直接打开官网反馈表单（不用注册账号）和 GitHub 讨论区；App Store 版还能一键去 App Store 评价
• 应用菜单改为自建，替换掉 Electron 默认那套指向 Electron 文档的「帮助」菜单；菜单文案跟随界面语言（32 种）
• 「关于」里显示真实版本号（之前写死为 0.1.0）
• 应用只允许打开官网、仓库与 App Store 评价链接，其它外部地址一律拒绝

Feedback made easy

• The Help menu and Settings → About now open the website feedback form (no account needed) and GitHub Discussions; the App Store build can also jump straight to "Rate on the App Store"
• The app now builds its own menu instead of Electron's default Help menu, which pointed at Electron's own docs; menu labels follow the interface language (32 languages)
• About shows the real version number (it was hard-coded to 0.1.0)
• The app only opens its own website, repository and App Store review links; any other external URL is refused

---

下载 / Downloads

	文件 / File	说明 / Notes
macOS	Wave.Subs-1.0.2-arm64.dmg	Apple Silicon，已公证 / notarized
macOS	Wave.Subs-1.0.2-arm64-mac.zip	同上，ZIP / same, as ZIP
Windows	Wave.Subs.Setup.1.0.2.exe	x64 安装程序（暂未签名，SmartScreen 会提示「更多信息 → 仍要运行」）/ x64 installer, not yet code-signed
Windows	Wave.Subs-1.0.2-win.zip	便携版 / portable

校验值见 SHA256SUMS.txt。也可以用包管理器：brew install --cask jason-jm/wavesubs/wavesubs · scoop bucket add wavesubs https://github.com/jason-jm/scoop-wavesubs && scoop install wavesubs

Checksums in SHA256SUMS.txt. Package managers: brew install --cask jason-jm/wavesubs/wavesubs · scoop bucket add wavesubs https://github.com/jason-jm/scoop-wavesubs && scoop install wavesubs

Wave Subs 1.0.1

看片找不到字幕？Wave Subs 用本地 AI 从影片生成 SRT / ASS 字幕，并自动翻译到你的语言。无需联网，全免费。

1.0.1

• 翻译目标语言从 5 种扩展到 29 种（法、德、西、葡、俄、泰、越、印尼、马来等），语言名按界面语言显示
• 首次启动默认翻译到系统语言，不再一律中文
• 界面颗粒纹理默认从 0.5 降到 0.1，不再有明显磨砂感；已装用户若从未调过会自动改为新默认
• 沙盒环境下源目录不可写时，字幕自动写到「影片 / Wave Subs」并在界面提示（App Store 版）

下载

平台	文件
macOS 12+（Apple Silicon，已公证）	Wave.Subs-1.0.1-arm64.dmg（或 -mac.zip）
Windows 10+（x64）	Wave.Subs.Setup.1.0.1.exe（或便携版 -win.zip）

校验值见 SHA256SUMS.txt。Windows 版尚未代码签名，SmartScreen 提示时点「更多信息 → 仍要运行」。

---

Can't find subtitles? Wave Subs generates SRT / ASS subtitles from any video with local AI and auto-translates them into your language. No internet needed, free forever.

1.0.1: 29 translation target languages (up from 5), first launch defaults to your system language, a subtler interface texture, and a sandbox-safe output fallback for the App Store build.

Wave Subs 1.0.0

Wave Subs 1.0.0 — 首个公开版本 / initial public release.

下载 / Downloads

• macOS（Apple Silicon，macOS 12+）：Wave.Subs-1.0.0-arm64.dmg（已由 Apple 公证）
• Windows（x64，Windows 10+）：Wave.Subs.Setup.1.0.0.exe 安装程序 / Wave.Subs-1.0.0-win.zip 便携版（未签名，SmartScreen 会提示「更多信息 → 仍要运行」）

功能 / What's inside

• 本地语音识别（whisper.cpp，Metal 加速）+ 时间轴精修（VAD + 响度，按六部整片基准校准）
• 三种字幕来源：识别 / 内嵌字幕轨（按语言自动选轨）/ 18 种外部字幕格式
• 翻译：本地 Qwen3 或任意 OpenAI 兼容接口；术语表
• 批量队列（整季一起跑，单文件可单独设置）
• 编辑器 + 视频预览（HEVC/DTS 等格式靠随包 ffmpeg 解码）
• 质检报告：覆盖率、漏段、超长、语速、缺译、残留原文
• 阶段缓存：换导出格式/重试秒级完成；换模型或术语表严格重翻
• 32 种界面语言

隐私 / Privacy：没有账号、没有统计、没有服务器；视频从不上传。只有你自己配置云端翻译时，字幕文本才会发给你选的服务商。

官网 / Website: https://jason-jm.github.io/wavesubs/
第三方组件许可 / Third-party licenses: THIRD-PARTY-LICENSES.md