mai2srt v0.1.4

mai2srt v0.1.4

本次更新修复两处影响长音频转录的问题，升级后长录音可以一次转录完成：自动分块、自动拼接，无需干预。

长音频转录被拒绝（HTTP 403）

• 现象：长录音（约 50 分钟左右起）转录报 HTTP 403，重试无效；更短的音频不受影响。
• 修复：自动分块的时长上限从 50 分钟降到 29 分钟。长音频切成更短的块分别转录后自动按时间戳拼接，输出不变；分块变短后每块码率也随之提高，转录质量更好。30 分钟以内的文件仍然单块直传，不受影响。

长视频准备阶段慢且无进度

• 现象：视频文件在切块前的静音检测会连同画面一起解码，视频越长等得越久；准备期间界面没有任何反馈，看起来像卡死。
• 修复：静音检测只解码音频流，速度大幅提升；准备过程的每一步（静音检测、切块规划、逐块编码）实时显示在任务日志与阶段指示上。

安装

下载下面的 mai2srt_0.1.4_x64-setup.exe 安装，FFmpeg 已随包内置。可以直接覆盖安装。

首次安装可能出现 Windows SmartScreen 蓝窗（安装包未签名）→ 点「更多信息」→「仍要运行」。附件中的 SHA256SUMS.txt 可用于校验下载是否完整。

说明

转录走 Microsoft AI Playground 的免费预览服务，请遵守其服务条款——它是预览服务，随时可能变动或下线。本工具为个人学习与低频使用编写，请自行评估使用风险。许可证：GPL-3.0。

---

English

This release fixes two problems affecting long-audio transcription: after upgrading, long recordings transcribe in one go — automatic chunking, automatic stitching, no manual steps.

Long-audio transcriptions rejected (HTTP 403)

• Symptom: long recordings (roughly 50 minutes and up) fail with HTTP 403; retries don't help. Shorter files are unaffected.
• Fix: the automatic chunking cap drops from 50 to 29 minutes. Long audio is split into shorter chunks, transcribed separately, and stitched automatically by timestamp; the output is unchanged. Shorter chunks also raise the per-chunk bitrate, improving transcription quality. Files under 30 minutes still upload as a single piece, unaffected.

Slow, silent video preparation

• Symptom: the pre-chunking silence scan decoded the video's picture stream along with the audio, so longer videos waited minutes in the prepare stage with no UI feedback — it looked like a hang.
• Fix: the silence scan decodes the audio stream only (much faster), and every preparation step (silence scan, chunk plan, per-chunk encode) now shows in real time in the job log and stage indicator.

Install

Download mai2srt_0.1.4_x64-setup.exe below — FFmpeg is bundled. It installs over an existing copy. Windows SmartScreen may warn on first install (unsigned installer): "More info" → "Run anyway". Verify the download against SHA256SUMS.txt.

Notes

Transcription runs on Microsoft AI Playground's free preview service — please respect its terms of service. Written for personal study and low-frequency use; judge the risks yourself. License: GPL-3.0.

Full Changelog: v0.1.3...v0.1.4

mai2srt v0.1.3

mai2srt v0.1.3

修复转录因账号未接受「生物特征通知」而被 HTTP 451 拒绝的问题（#1）：playground 把音频视为生物特征数据（BIPA），账号须先接受一次通知才能上传音频。本次更新解释了这个报错并给出明确的指引，在桌面端与命令行各加了一个接受入口，并让新账号在登录时就完成检测——已经接受过的账号不受任何影响。

生物特征通知

• playground 把音频视为生物特征数据（BIPA）：账号需要先接受一次「生物特征通知」才能上传音频。网页端靠首次上传时弹出的同意框完成这一步，而本工具走无头通道看不到弹窗，所以需要自己的接受入口。
• 新账号登录完成时会自动检测：需要接受的话，设置页的「接受生物特征通知」按钮会高亮并给出提示，点一次即可。
• 桌面端：设置 → 文件与账号 → 「接受生物特征通知」；已经撞上 451 的话，转录错误信息也会直接指引到这里。
• 命令行：新账号 mai2srt login 完成后会当场询问，确认即接受；转录遇到 451 时也会当场询问，确认后自动接受并继续，无需重跑。也可以随时运行 mai2srt consent。
• 接受记录保存在微软账号上，一次生效、长期有效，之后转录正常进行。
• 本工具不会替你静默接受：这是法律意义上的同意，确认永远由你本人做出。多账号用户请注意，每个账号需要各自接受一次。

安装

下载下面的 mai2srt_0.1.3_x64-setup.exe 安装，FFmpeg 已随包内置。可以直接覆盖安装。

首次安装可能出现 Windows SmartScreen 蓝窗（安装包未签名）→ 点「更多信息」→「仍要运行」。附件中的 SHA256SUMS.txt 可用于校验下载是否完整。

说明

转录走 Microsoft AI Playground 的免费预览服务，请遵守其服务条款——它是预览服务，随时可能变动或下线。本工具为个人学习与低频使用编写，请自行评估使用风险。许可证：GPL-3.0。

---

English

Fixes transcriptions being rejected with HTTP 451 when the account has not yet accepted the biometric notice (#1): playground treats audio as biometric data (BIPA), and each account must accept the notice once before audio uploads. This release explains the error and provides clear guidance, adds an acceptance entry point in both the GUI and the CLI, and checks fresh accounts at sign-in — accounts that have already accepted are unaffected.

Biometric notice

• playground treats audio as biometric data (BIPA): an account must accept the "biometric notice" once before uploading audio. The website handles that with a consent dialog on the first upload; mai2srt's headless channel never sees the dialog, so the acceptance needs its own entry points here.
• A fresh sign-in checks automatically: when the notice is still owed, the "accept biometric notice" button in Settings is highlighted with a hint — one click and it is done.
• GUI: Settings → Files & accounts → "accept biometric notice"; if a run hits the 451, the transcription error itself points there.
• CLI: mai2srt login on a fresh account asks right after sign-in — confirm once and it is done; a transcription that hits the 451 asks on the spot too, and continues without a restart after confirming. mai2srt consent works at any time.
• The acceptance is stored on your Microsoft account, once and for all; transcription works normally afterwards.
• mai2srt never accepts silently on your behalf: this is a legal consent, and the confirmation is always yours to make. With multiple accounts, each account needs its own one-time acceptance.

Install

Download mai2srt_0.1.3_x64-setup.exe below — FFmpeg is bundled. It installs over an existing copy. Windows SmartScreen may warn on first install (unsigned installer): "More info" → "Run anyway". Verify the download against SHA256SUMS.txt.

Notes

Transcription runs on Microsoft AI Playground's free preview service — please respect its terms of service. Written for personal study and low-frequency use; judge the risks yourself. License: GPL-3.0.

Full Changelog: v0.1.2...v0.1.3

mai2srt v0.1.2

mai2srt v0.1.2

针对大模型设置的思考档位做了修复：选定的档位会按各家支持的方式正确发送，模型列表也能识别出 Claude 和 Gemini 的可思考模型。另外，添加供应商时只显示该家实际支持的 API 格式。

思考强度档位

• 档位按各家自己的方式发送：新版 Claude 用 effort 参数，旧版 Claude 和 Gemini 2.5 用思考预算，Gemini 3 用思考档位。
• 模型不认收到的参数形式时，会自动换用它接受的形式重试；实在不支持就关闭思考继续完成断句，不会整批退回规则切分。
• 「测试」按钮按真实作业的请求方式发送：选中思考档位时，测试请求也会带上思考参数，测试通过即代表实际断句请求同样可用。
• Gemini 3 的回复上限包含思考部分，会自动留出余量，避免长回复因思考占用而截断。
• 选 off 时，Gemini 2.5 会真正关闭思考；Gemini 3 本身无法关闭思考，保持默认。

模型列表

• Claude（3.7 及以上）与 Gemini（2.5 及以上）会正确显示可思考图标。
• 服务商自定义的模型名（如 gemini-3-flash）不提供数值上限时，会使用内置的兜底数值。

添加供应商

• 每个预设只显示该家实际支持的 API 格式：DeepSeek、OpenRouter、Ollama（本地）支持 OpenAI 和 Anthropic 两种；OpenAI 官方和 Anthropic 官方各自只有一种；Gemini 支持原生和 OpenAI 兼容两种；自定义仍可选全部三种。
• 切换格式时，接口地址会自动跟随更换；手动修改过的地址不受影响。
• 使用 Ollama 的 Anthropic 格式需要 v0.14.0 及以上版本。

安装

下载下面的 mai2srt_0.1.2_x64-setup.exe 安装，FFmpeg 已随包内置。可以直接覆盖安装 0.1.0 与 0.1.1。

首次安装可能出现 Windows SmartScreen 蓝窗（安装包未签名）→ 点「更多信息」→「仍要运行」。附件中的 SHA256SUMS.txt 可用于校验下载是否完整。

说明

转录走 Microsoft AI Playground 的免费预览服务，请遵守其服务条款——它是预览服务，随时可能变动或下线。本工具为个人学习与低频使用编写，请自行评估使用风险。许可证：GPL-3.0。

---

English

The thinking-level setting for LLM providers is fixed: the level you pick is now sent in the form each provider accepts, and the model list recognises thinking-capable Claude and Gemini models. The add-provider card also only shows the API formats a vendor really serves.

Effort selector

• The selected level is sent the way each vendor expects: newer Claude takes the effort parameter, older Claude and Gemini 2.5 take a token budget, Gemini 3 takes a thinking level.
• When a model rejects the parameter form it receives, the splitter retries with one it accepts; if thinking is unsupported it is turned off and the job completes, without the whole batch falling back to rule-based splitting.
• The "test" button sends the same request shape as a real job: with a thinking level selected, the test request carries the thinking parameter too, so a passing test covers the real request.
• Gemini 3 counts thoughts against the reply limit; a margin is reserved so long replies are not truncated.
• "off" genuinely disables thinking on Gemini 2.5; Gemini 3 cannot turn thinking off, so the default stays.

Model list

• Claude (3.7 and up) and Gemini (2.5 and up) now show the thinking badge.
• Gemini model names that report no numeric limits (e.g. custom aliases like gemini-3-flash) get built-in fallback values.

Adding a provider

• Each preset only shows the formats the vendor actually serves: DeepSeek, OpenRouter and Ollama (local) offer OpenAI and Anthropic; the OpenAI and Anthropic presets offer their own single format; Gemini offers native and OpenAI-compatible; custom still offers all three.
• Switching the format retargets the endpoint URL; a hand-edited URL is left alone.
• Ollama's Anthropic format requires v0.14.0 or newer.

Install

Download mai2srt_0.1.2_x64-setup.exe below — FFmpeg is bundled. It installs over 0.1.0 and 0.1.1. Windows SmartScreen may warn on first install (unsigned installer): "More info" → "Run anyway". Verify the download against SHA256SUMS.txt.

Notes

Transcription runs on Microsoft AI Playground's free preview service — please respect its terms of service. Written for personal study and low-frequency use; judge the risks yourself. License: GPL-3.0.

Full Changelog: v0.1.1...v0.1.2

mai2srt v0.1.1

mai2srt v0.1.1

精修页的拆分栏重做，另外修掉一批"中间状态"上的交互问题，以及提升了断句能力。

精修页拆分

• 拆分栏展开后分成两行：上面选精度，下面才是可点的字块。
• 新增「句 / 字」两档精度，默认按句：剪刀只出现在标点位置，不用再自己在一排字里找逗号。
• 切到按字时，标点字块高亮成实心主题色，句末与句中停顿一视同仁——一眼就能看到该切的地方。
• 点该行以外的任何位置都会退出当前状态；正在编辑的文字会自动保存，想放弃就按 Esc 或点「取消」。

修好的问题

• 删除待确认、或拆分栏开着时，点合并按钮会直接完成合并、丢掉那个待确认状态；现在拆分与删除互斥，中间状态下只有铅笔图标始终可点。
• 这两种状态下还能点单句试听，导致同一行上两个状态叠着；现在会锁住。
• 点其他行的空白处或时间戳区域不会退出当前状态；现在会退出（并把编辑保存下来）。
• 转录页首次打开时，状态灯与参数条可能不显示，要切到别的页面再切回来才出现；现在正常。
• 长录音的断句上下文摘要会因超长而静默失效；现在按预算裁剪，并在日志里说明。
• 在精修页拆分行时，中文标点前会多出一个空格；现在不会。

大模型断句

• 分批依据改成模型自身的能力上限。
• 逐批显示进度，结束时报告有多少段由大模型给出。
• 个别段落没能对齐时会明确提示。
• 设置页的模型徽章同时显示上下文与最大输出；由名称推断的数值带 ≈ 前缀，服务商自己上报的不带。

开始转录前的检查

• 点开始前会先确认 ffmpeg、浏览器、playground 登录，以及勾选大模型断句时的大模型配置；缺哪一项会提示去哪配，不再等作业跑起来才失败。
• 只处理 .mai.json 时只检查大模型配置。

安装

下载下面的 mai2srt_0.1.1_x64-setup.exe 安装，FFmpeg 已随包内置。可以直接覆盖安装 0.1.0。

首次安装可能出现 Windows SmartScreen 蓝窗（安装包未签名）→ 点「更多信息」→「仍要运行」。附件中的 SHA256SUMS.txt 可用于校验下载是否完整。

说明

转录走 Microsoft AI Playground 的免费预览服务，请遵守其服务条款——它是预览服务，随时可能变动或下线。本工具为个人学习与低频使用编写，请自行评估使用风险。许可证：GPL-3.0。

---

English

The Refine page's split bar is rebuilt, a batch of interaction problems around the "in between" states is fixed, and segmentation got better.

Splitting in the Refine page

• The bar now opens as two rows: pick the granularity, then the chips below it.
• New "sentence / character" choice, sentence by default: scissors appear only at punctuation, so you no longer hunt for a comma among a row of characters.
• In character mode punctuation chips are filled with the accent colour — sentence ends and mid-sentence pauses alike — so the place you want to cut is obvious.
• Clicking anywhere outside that line leaves the current state; text being edited is saved automatically (press Esc or use "cancel" to discard).

Fixed

• With a delete armed or the split panel open, clicking a merge button merged the lines and silently dropped that pending state; split and delete are now mutually exclusive, and only the pencil stays clickable in every state.
• Those two states also allowed playing a single cue, leaving two live states on one line; now locked.
• Clicking another line's blank space or timestamp area did nothing; it now leaves the current state (saving any edit).
• On the Transcribe page the status lights and the params bar could be missing until you switched pages and back; fixed.
• The LLM's whole-job context summary failed silently on long recordings; it is now trimmed to the budget and says so in the log.
• Splitting a line inserted a stray space before Chinese punctuation; fixed.

LLM segmentation

• Batching now follows the model's own limits.
• Per-batch progress, and a closing count of how many segments the model actually produced.
• Segments the model failed to align are now reported.
• Model badges in Settings show both context and max output; values inferred from the name carry a ≈ prefix, values the provider reports do not.

Checks before a run starts

• Starting a transcription first verifies ffmpeg, a browser, the playground sign-in, and — when LLM splitting is on — the LLM endpoint. Whatever is missing is named, instead of failing once the job is already running.
• A .mai.json run only checks the LLM endpoint.

Install

Download mai2srt_0.1.1_x64-setup.exe below — FFmpeg is bundled. It installs over 0.1.0. Windows SmartScreen may warn on first install (unsigned installer): "More info" → "Run anyway". Verify the download against SHA256SUMS.txt.

Notes

Transcription runs on Microsoft AI Playground's free preview service — please respect its terms of service. Written for personal study and low-frequency use; judge the risks yourself. License: GPL-3.0.

mai2srt v0.1.0

mai2srt v0.1.0

第一个公开版本：把音频或视频拖进窗口，得到词级时间戳的转录，按自己的习惯精修断句，导出标准 SRT。

能做什么

• 常见音频与视频格式直接拖入；超过 playground 限制（25 MiB / 约 60 分钟）的文件自动压缩、按静音分块，词级时间戳拼回完整时间轴
• 转录区分出的说话人会保留；几人挨得太近分不开时并成一条多行字幕
• 断句不满意就在精修页逐行改：拆分 / 合并 / 删除、单句试听、点行跳播、撤销重做，编辑记录随项目保存
• 断句可以让大模型帮忙挑位置（内置 DeepSeek / OpenAI / Claude / Gemini / OpenRouter / Ollama 预设，也可填任意兼容端点）；不配也能出字幕，大模型不改文字、不动时间戳
• 玫瑰（明亮）/ 余烬（暗黑）两套主题，各配液态 / 毛玻璃材质与自定义壁纸；「跟随系统」按本机时间走——19:00 之后转暗，07:00 亮回来
• 命令行 mai2srt（从源码安装后可用）：transcribe / process / run，一次一个文件，适合写进脚本；process 纯离线重跑断句，不消耗转录额度

安装

下载下面的 mai2srt_0.1.0_x64-setup.exe 安装即可，FFmpeg 已随包内置，不需要另装 Python 或 FFmpeg。

首次安装可能出现 Windows SmartScreen 蓝窗（安装包未签名）→ 点「更多信息」→「仍要运行」。装完第一次启动会慢一点：后端进程首次拉起，加上杀毒软件扫描新文件；之后启动很快。

校验

附件中的 SHA256SUMS.txt 可用于校验安装包下载是否完整。

说明

转录走 Microsoft AI Playground 的免费预览服务，请遵守其服务条款——它是预览服务，随时可能变动或下线。本工具为个人学习与低频使用编写，请自行评估使用风险。许可证：GPL-3.0。

---

English

First public release. Drop in audio or video, get a word-level transcript, refine the segmentation the way you like it, export a standard SRT.

• Common audio and video formats drop straight in; files over the playground limits (25 MiB / ~60 min) are compressed and split on silence, and the word timestamps are stitched back into one timeline
• Speakers are kept apart; when several people talk too close together to separate, their words are folded into one multi-line subtitle
• The Refine page edits the result line by line — split / merge / delete, listen to a single cue, click a line to jump, undo and redo, with the edit record kept with the project
• Segmentation can lean on an LLM (presets for DeepSeek / OpenAI / Claude / Gemini / OpenRouter / Ollama, or any compatible endpoint) or stay rule-based; the LLM never rewrites text or timestamps
• Two themes, rose (light) and ember (dark), with liquid / frosted glass and custom wallpapers; "follow the system" runs on the clock — dark after 19:00, light again at 07:00
• A mai2srt CLI (from a source install): transcribe / process / run, one file per run, scriptable; process re-runs segmentation fully offline without spending transcription quota

Install with mai2srt_0.1.0_x64-setup.exe below — FFmpeg is bundled, no Python or FFmpeg to install. Windows SmartScreen may warn on first install (unsigned installer): "More info" → "Run anyway". The first launch takes a moment while the backend starts and antivirus scans the new files. Verify the download against SHA256SUMS.txt.

Transcription runs on Microsoft AI Playground's free preview service — please respect its terms of service. Written for personal study and low-frequency use; judge the risks yourself. License: GPL-3.0.

Full Changelog: https://github.com/fuxiaomoke/mai2srt/commits/v0.1.0