言流 YanFlow

让声音落在光标上，而不是云端。
Let your voice land on the cursor—not in the cloud.

下载免安装包 / Download · 中文 · English

Video / 视频 — The moving picture belongs here. Demo coming soon. 影像证词，稍后抵达。

中文

言流是 Windows 上的全局实时语音转录工具。默认按住 Ctrl + Win 录音，松开识别，结果会 落进此刻正在编辑的地方；没有合适的输入框时，它就把文字留在一枚可复制、可写入 TXT 的气泡里。

纯本地处理，纯 CPU 完成端点检测和识别。不用担心隐私问题，不需要账号、网络、Python、 PyTorch，更不要求独立显卡。模型由常驻 worker 只加载一次，采集回调只写固定无锁环形 缓冲；没有临时 WAV，没有每句话一次冷启动。识别队列有上限，悬浮窗只在状态变化时重绘。

它会什么

• 全局落字：记事本、浏览器地址栏、VS Code，以及其他标准 Edit/Document 控件。
• 不抢焦点：原生 Win32 无激活悬浮窗；拖动、隐藏、复制、快捷键设置都在手边。
• 离线端侧：16 kHz 单声道采集，自适应噪声底、300 ms 预卷、最长 8 秒连续切片。
• 保护切片边界：长录音优先在静音附近切分；连续语音使用小幅重叠和 CTC 时间对齐，单次推理仍不超过 8 秒。
• 短词也提交：按住模式单独支持短词，经过较严格的人声复核；静音和短促敲击仍被过滤。
• 安静时停下：30 秒没有确认的人声后关闭麦克风；边界上的一句话最多再等 8 秒完成识别。
• 减少误触发：每段录音经过 FSMN-VAD 复核；静音和短促敲击不会直接送入转录。高通滤波减轻低频风噪。
• 失焦也不失语：目标不可写时，转录进入向左展开的气泡，可一键复制或写入桌面。
• 开箱即用：下载 Release ZIP，解压，双击 yanflow.exe。没有安装器，也没有“下一步”。

三步，声音就有了去处

1. 从 Releases 下载 windows-x64.zip。
2. 解压到任意可写目录，运行 yanflow.exe。
3. 按住 Ctrl + Win 录音，松开任一键识别。免按住模式仍用 Ctrl + Alt + Space 开始、 Ctrl + Alt + S 停止；右键悬浮按钮可以改键或关闭按住模式。按住模式最长 60 秒。

保守纠错

词典先做明确的大小写规范和术语映射；音近词必须列出误识别形式和同句上下文，默认不做 模糊猜测。右键选择“编辑自定义词典”，保存后下一次识别生效。MacBERT4CSC 是默认关闭的 可选本地 C++ worker，只接受高置信度中文单字替换，并保护词典术语、英文、数字和可识别的代码片段。 资源缺失、超时或异常时保留第一阶段结果。基础包仍不需要 Python 或额外运行时安装。

右键“保存最近识别对照”可导出原始 ASR 和最终文本，人工校对后评测 CER 和误改。 右键“纠正并记住”可确认误词→正确词，默认只用于这次录音的目标应用，并复制本次正确文本。 “应用词库”可选择自动、通用或开发词库；应用专用词库叠加在所选词库上，升级不会覆盖个人规则。 启用可选模型、词典格式及三阶段评测方法见 纠错说明。

Windows 可能在首次使用麦克风时请求权限。普通权限进程不能向管理员权限窗口注入文本， 这是 Windows UIPI 的边界，不是言流在故作矜持。

English

Some words should not have to leave the room before they become text.

YanFlow is a system-wide voice layer for Windows. It floats above every app without stealing focus. Hold Ctrl + Win to record and release to recognize; the transcript lands in whatever you are editing. If there is no writable target, the words wait in a compact bubble—ready to copy or save as TXT.

Its “cloud” is the few centimetres of air above your desk. SenseVoiceSmall Q8 and FSMN-VAD run entirely on the local CPU: no audio upload, account, network service, Python, PyTorch, or discrete GPU. A resident worker loads the model once; audio capture writes into a fixed lock-free ring buffer. No temporary WAV files. No cold start for every sentence. The footprint does not vanish by incantation—it is simply kept inside deliberate boundaries.

What it does

• Types system-wide into Notepad, browser address bars, VS Code, and other Edit/Document controls.
• Keeps your focus with a native non-activating Win32 overlay and configurable global hotkeys.
• Stays on-device with 16 kHz mono capture, adaptive noise floor, 300 ms pre-roll, and 8 s slices.
• Stops an idle microphone after 30 s without confirmed speech, with one brief grace period for a phrase in progress.
• Checks speech with FSMN-VAD before transcription and attenuates low-frequency ventilation rumble.
• Catches stray words in a left-expanding bubble when the target cannot accept text.
• Ships ready as a portable ZIP: extract and run yanflow.exe; there is nothing to install.

Three steps from air to cursor

1. Download the windows-x64.zip from Releases.
2. Extract it to any writable directory and run yanflow.exe.
3. Hold Ctrl + Win to record; release either key to recognize (60 s maximum). Hands-free mode still starts with Ctrl + Alt + Space and stops with Ctrl + Alt + S. Right-click the orb to remap hands-free keys or disable hold mode.

An editable dictionary normalizes explicit terms conservatively. Optional MacBERT4CSC is disabled by default and only accepts tightly gated Chinese character edits; dictionary terms, English, numbers, and recognizable code are protected. Export the last raw/final pair from the orb menu to measure improvements against a human transcript. See correction details.

Windows may request microphone permission on first use. A normal process cannot inject text into an elevated process because of UIPI; that boundary is intentional.

Build / 构建

YanFlow is a native Windows x64 application written in C++17, using Win32 for the GUI, audio capture, hotkeys, and text injection. Both the GUI and recognition worker statically link the MSVC runtime; PHP, TypePHP, and a separate VC redistributable are not required. On Windows, install Visual Studio 2022 Build Tools with the Desktop development with C++ workload, including the Windows SDK and C++ CMake tools (CMake and Ninja):

build\windows\build-yanflow.cmd

The script downloads pinned SenseVoice/FSMN resources, verifies SHA-256 values, builds the persistent C++ worker and native GUI with CMake, embeds the icons as Windows resources, then stages a fresh build\artifacts\yanflow-windows-x64\ and runs window and ASR smoke tests. VS_BUILD_TOOLS remains overridable. The previous TypePHP implementation is preserved on the history/typephp-2026-10-08 branch.

Pushes and pull requests build that same portable package on GitHub Actions. A v* tag also creates a GitHub Release with the ZIP and SHA256SUMS automatically.

Useful focused checks / 常用专项验证：

build\windows\test-yanflow-worker.cmd
build\windows\test-yanflow-native.cmd
build\windows\test-yanflow-idle.cmd
build\windows\test-yanflow-pipeline.cmd
build\windows\test-yanflow-textbox.cmd
build\windows\test-yanflow-fallback.cmd
build\windows\test-yanflow-edge.cmd
build\windows\test-yanflow-vscode.cmd
build\windows\test-yanflow-hold.cmd
build\windows\test-yanflow-correction.cmd
build\windows\test-yanflow-accuracy.cmd

The native-package check verifies PE imports, creates build\artifacts\YanFlow-native-windows-x64.zip and its SHA256SUMS-native.txt, then extracts and runs the application from a path containing spaces and Unicode. It also checks the missing-wave and missing-worker exit codes. When the optional CSC bundle is present, it verifies its app-local DLL dependencies and runs the actual model from the extracted path. To build that bundle, run build\windows\setup-yanflow-csc.cmd, then build\windows\build-yanflow.cmd; ONNX Runtime and the model are pinned and SHA-256 checked.

Boundaries / 边界

Today YanFlow targets Windows x64 and CPU inference. Automated evidence covers persistent-worker repetition, the full segmentation pipeline, standard text boxes, Edge, VS Code, and the fallback bubble. A combined live-human microphone session and a 24–72 hour soak remain future evidence—not retroactive folklore. Hold-mode automated evidence uses injected PCM and a key-state smoke. Physical modifier chords, Start-menu masking, session lock/resume, and a live human hold-to-record session require manual Windows validation. Synthetic correction fixtures establish guard behavior, not a lower live-ASR error rate; that requires paired real speech and human transcripts. Accuracy smoke covers quiet and forced boundaries in repeated 19.2 s PCM fixtures, a 300 ms speech fragment through capture/flush/VAD/ASR, short silence/tap rejection, and the native correction dialog with dictionary persistence and application isolation. CTC times are approximate; isolated syllables can still be misrecognized. These fixtures do not establish live microphone accuracy or overall CER gains.

Single-microphone filtering and VAD can reduce steady fan noise, silence, and brief impacts. They cannot reliably isolate a nearby speaker from background voices or cancel room echo without a playback reference. CPU load and heat also depend on microphone input and the host machine; the automated worker smoke checks short-run working-set drift, not long-term thermal behavior.

YanFlow is GPL-3.0-only. Models and bundled runtimes retain their own licenses; see third-party notices and the license files included in every release ZIP.