在线模式和离线模式有什么区别？

在线模式是实时流式：一边采集麦克风一边出字，可选用本地 WASM 引擎（断网可用）或浏览器云端识别服务（需联网，音频会发给浏览器厂商）。离线模式是文件转写：选择本地音频/视频文件后本地解码、本地推理，不依赖麦克风。

输出的 who / when / what 三块是怎么来的？

模型按 JSON 输出时（Start/End/Speaker/Content），说话人与时间戳由模型直接给出；如果模型只给了纯文本，就按句切分、用本段音频区间按字数比例分配时间，说话人标记为「说话人 1」。数据层会强制补齐缺失项：没有结束时间就按起始 + 2 秒兜底，内容为空显示「（无内容）」，因此每块必然三块齐全且顺序固定为 who → when → what。

实时流式是怎么做到「边说边出字」的？

按窗口分块流式：累积到设定窗口（默认 5 秒）或检测到说话停顿时，把这一段送进引擎，结果与上一段做重叠去重后追加。它不是逐 token 流式，但停顿即可出字，延迟就是窗口大小。

3A 是什么，为什么有些开关没生效？

3A 指 AEC 回声消除、ANS 噪声抑制、AGC 自动增益，由浏览器内置的 WebRTC 音频处理模块提供，本工具通过 getUserMedia 约束开关并回读 getSettings()。是否生效取决于浏览器与设备，面板会显示实际生效值；不支持的扩展项（人声隔离、AEC 模式）会自动跳过。

为什么第一次离线转写很慢？

首次需要下载两个 GGUF 模型（VAE 编码器约 0.65 GB、LM 解码器约 0.92 GB，共约 1.6 GB），页面会显示实时下载进度；下载完成后写入浏览器缓存，之后显示「已缓存（可离线使用）」并不再联网。

离线识别用的是什么引擎？

微软官方的 VibeASR.cpp（依赖 llama.cpp）用 Emscripten 编译成 WASM，仓库里 wasm/ 目录是封装代码，执行 npm run build:wasm 可自行构建；推理跑在 Web Worker 中，不会卡住页面。

浏览器跑 1.6GB 模型会有问题吗？

模型权重会进入 WASM 内存，建议 8 GB 以上内存的桌面浏览器使用，并保持页面在前台；首次装载需要几十秒，装载完成后每次转写只处理音频本身。

识别结果不准怎么办？

安静环境、距麦克风 20~40 厘米；开启 ANS 抑制底噪、AGC 拉齐音量；长音频分段处理；语言选择与实际语种一致时准确率更高。

语音转文字

在线实时流式与离线文件转写：在线模式麦克风边说边出字，离线模式选择音频/视频文件本地转写并保留时间轴；本地 WASM 跑 VibeVoice-ASR-BitNet，模型可联网下载（官方源连不上自动改用 hf-mirror.com 镜像）、可指定保存目录、也可直接装载本地文件，支持 SHA-256 完整性校验与断网使用；采集链路支持 WebRTC 3A（回声消除/降噪/自动增益），可导出 TXT/SRT。

语音转文字asr语音识别实时流式wasm离线vibevoice降噪3a字幕hf-mirror镜像sha256speech to text

关于「语音转文字」

它属于「AI 本地推理」分类：用 WASM 在浏览器本地跑模型：语音转文字等能力不联网也能用，音视频不上传服务器。

在线实时流式：麦克风音频边录边送进识别引擎，文字随说话实时追加，适合会议记录、实时字幕。
离线文件转写：选择或拖入音频/视频文件，本地解码后自动语音分段、逐段转写并保留时间轴，可导出 TXT 与 SRT。
离线识别跑 VibeASR.cpp（VibeASR.cpp + llama.cpp 编译的 WASM）：VAE 音频编码器 + LM 解码器两个 GGUF 模型，首次下载约 1.6 GB，之后完全离线、音频不出浏览器。
模型文件先查本地缓存，命中就直接装载；未命中才下载，下载后写入浏览器缓存，之后断网可用。
也支持「本地装载」：直接选择磁盘上的两个 GGUF 文件（可拖入或点选），全程不联网，适合内网或已经提前下好模型的场景；Chrome/Edge 上可勾选「记住这批文件」，下次打开页面授权一次即可复用。
联网下载时可以指定「保存目录」，模型会直接写进你选的那个文件夹（支持断点续传），之后从该目录读取，不再重复下载；目录里的模型是普通文件，能拷走、也能提前用其它工具下载好放进去。
官方源 huggingface.co 在国内经常连不上：页面会自动改用镜像 hf-mirror.com 重试，模型地址是固定的，不需要你填。哪个源成功过会被记住，下次优先使用；也可以用「下载源」下拉锁定单一源，或点「恢复官方优先」清掉记忆。若两个源都连不上（被墙，或被网络中间层剥掉了跨域响应头），面板会在「下载源」下方给出这两个文件的直达下载链接——点开即由浏览器直接下载，放进「保存目录」就能装载。
本地引擎实测速度：13.5 秒语音约需 187 秒（约 14× 实时，2 秒分块 + 单线程标量），装载 1.6GB 权重只要 2 秒级。所以实时连续识别请用「浏览器云端」引擎；本地引擎适合短音频与文件转写，长时间录音要有等待预期。开启 COOP/COEP 后可切到多线程版提速。
装载前会对每个模型文件做 SHA-256 校验：期望值优先取你手动填的哈希，其次是模型清单，再其次是官方仓库 API 与服务器 ETag。校验不通过会中止装载并给出「期望 vs 实际」的警告，你可以重新下载，也可以确认后选择忽略继续。拿不到期望值时会如实显示「未校验」，不会假装通过。
每条记录严格包含 who（对话人）、when（起止时间戳）、what（话术内容）三块：数据层统一规范化（缺结束时间自动补齐、空内容显式标注），渲染层每块固定三列，整列对齐。
瀑布流一块一行、垂直依次排列；导出同样保持三块——TXT 每行「说话人 | 00:03 - 00:07 | 内容」，SRT 带说话人，JSON 给出结构化字段。
采集链路可开启 WebRTC 3A：回声消除 AEC、噪声抑制 ANS、自动增益 AGC，并回读浏览器真正生效的参数。

完全免费、无需注册；所有处理都在浏览器本地完成，文件不会上传到任何服务器。

相关关键词：语音转文字、AI 本地推理、asr、语音识别、实时流式、wasm、离线、vibevoice、降噪、3a、字幕、hf-mirror、镜像、sha256、speech to text、在线工具、免费

如何使用「语音转文字」

打开工具：打开「语音转文字」页面，无需注册、无需安装任何软件。
准备输入：准备好要处理的内容或文件：文本可直接粘贴，文件可拖入页面（本工具属于「AI 本地推理」）。
填入内容：把内容粘贴进输入框，或把文件拖入页面；也可以点击「示例」先填入样例数据试一下。
设置选项并处理：按需调整页面上的选项，然后点击处理 / 转换 / 生成按钮。
获取结果：处理完全在浏览器本地完成，结果可直接复制或下载，文件不会上传到任何服务器。

常见问题

在线模式和离线模式有什么区别？: 在线模式是实时流式：一边采集麦克风一边出字，可选用本地 WASM 引擎（断网可用）或浏览器云端识别服务（需联网，音频会发给浏览器厂商）。离线模式是文件转写：选择本地音频/视频文件后本地解码、本地推理，不依赖麦克风。
输出的 who / when / what 三块是怎么来的？: 模型按 JSON 输出时（Start/End/Speaker/Content），说话人与时间戳由模型直接给出；如果模型只给了纯文本，就按句切分、用本段音频区间按字数比例分配时间，说话人标记为「说话人 1」。数据层会强制补齐缺失项：没有结束时间就按起始 + 2 秒兜底，内容为空显示「（无内容）」，因此每块必然三块齐全且顺序固定为 who → when → what。
实时流式是怎么做到「边说边出字」的？: 按窗口分块流式：累积到设定窗口（默认 5 秒）或检测到说话停顿时，把这一段送进引擎，结果与上一段做重叠去重后追加。它不是逐 token 流式，但停顿即可出字，延迟就是窗口大小。
3A 是什么，为什么有些开关没生效？: 3A 指 AEC 回声消除、ANS 噪声抑制、AGC 自动增益，由浏览器内置的 WebRTC 音频处理模块提供，本工具通过 getUserMedia 约束开关并回读 getSettings()。是否生效取决于浏览器与设备，面板会显示实际生效值；不支持的扩展项（人声隔离、AEC 模式）会自动跳过。
为什么第一次离线转写很慢？: 首次需要下载两个 GGUF 模型（VAE 编码器约 0.65 GB、LM 解码器约 0.92 GB，共约 1.6 GB），页面会显示实时下载进度；下载完成后写入浏览器缓存，之后显示「已缓存（可离线使用）」并不再联网。
离线识别用的是什么引擎？: 微软官方的 VibeASR.cpp（依赖 llama.cpp）用 Emscripten 编译成 WASM，仓库里 wasm/ 目录是封装代码，执行 npm run build:wasm 可自行构建；推理跑在 Web Worker 中，不会卡住页面。
浏览器跑 1.6GB 模型会有问题吗？: 模型权重会进入 WASM 内存，建议 8 GB 以上内存的桌面浏览器使用，并保持页面在前台；首次装载需要几十秒，装载完成后每次转写只处理音频本身。
识别结果不准怎么办？: 安静环境、距麦克风 20~40 厘米；开启 ANS 抑制底噪、AGC 拉齐音量；长音频分段处理；语言选择与实际语种一致时准确率更高。

Speech to Text (English)

Live streaming and file transcription: live mode turns microphone audio into text as you speak, file mode transcribes local audio/video with timestamps. Local WASM inference with VibeVoice-ASR-BitNet can be downloaded (auto-retried via the hf-mirror.com mirror), saved into a folder you choose, or loaded straight from local files — with SHA-256 integrity checks and no network needed afterwards; capture supports WebRTC 3A (AEC / noise suppression / AGC); export to TXT/SRT.

Belongs to “Local AI”: Run models locally in your browser with WASM: speech-to-text and more work offline, and audio/video is never uploaded.

Live streaming: microphone audio is fed to the engine continuously and text is appended as you speak — good for meeting notes and live captions.
File transcription: pick or drop an audio/video file; it is decoded locally, split by voice activity, transcribed segment by segment with timestamps, and exportable as TXT/SRT.
Offline recognition runs VibeASR.cpp compiled to WASM (with llama.cpp): a VAE audio encoder plus an LM decoder, two GGUF files about 1.6 GB in total. After the first download it works fully offline and audio never leaves the browser.
Model files are looked up in local cache first and loaded if present; only missing files are downloaded, then cached, so later sessions work offline.
You can also load models locally: pick the two GGUF files from disk (or drag them in) and nothing touches the network — handy for intranets or when you already have the files. On Chrome/Edge, tick "Remember these files" to reuse them next visit after a single permission prompt.
When downloading you can pick a save folder: models are written straight into it (resumable) and read from there afterwards, so they are never downloaded twice. The files are plain files — copy them around, or drop in models fetched with another tool.
huggingface.co is often unreachable from mainland China: downloads are retried through the hf-mirror.com mirror automatically, and the model URL is fixed so there is nothing to fill in. The source that worked is remembered and preferred next time; you can also lock one source in the dropdown, or click "Prefer official again" to clear that memory. If both sources fail (blocked, or a middlebox strips the cross-origin headers), the panel shows direct download links under "Download source" — clicking them downloads via the browser, and dropping the files into the save folder is all it takes.
Measured locally: 13.5s of speech takes ~187s to transcribe (~14× real-time, 2-second chunks, single-threaded scalar), while loading the 1.6GB weights takes only seconds. So use the browser-cloud engine for live recognition; the local engine suits short clips and file transcription, and long recordings need patience. Enabling COOP/COEP switches it to the multi-threaded build for a speed-up.
Every model file is checked with SHA-256 before loading: the expected value comes from your manual entry, then the manifest, then the official repo API or the server ETag. A mismatch stops loading and shows an expected-vs-actual warning — re-download, or confirm and continue anyway. When no expected value is available it says "not checked" rather than pretending to pass.
Every record strictly contains who (speaker), when (start–end timestamps) and what (the utterance): normalized in the data layer (missing end time is filled in, empty content is marked explicitly) and rendered as three aligned columns per block.
The waterfall lists one block per line, vertically; exports keep the same three fields — TXT lines as "speaker | 00:03 - 00:07 | content", SRT with speaker prefix, JSON with structured fields.
The capture chain supports WebRTC 3A: echo cancellation (AEC), noise suppression (ANS) and auto gain control (AGC), with the actually applied parameters read back from the browser.

Free, no sign-up; everything runs locally in your browser and files are never uploaded.

Keywords: Speech to Text, Local AI, speech to text, asr, voice, streaming, wasm, offline, vibevoice, 3a, noise suppression, subtitle, hf-mirror, sha256, transcribe, online tool, free

How to use it

Open the tool：Open the “Speech to Text” page — no sign-up and nothing to install.
Prepare the input：Prepare the content or file to process: paste text or drop a file onto the page (this tool belongs to “Local AI”).
Provide the input：Paste your content into the input box or drop a file onto the page; you can also click “Sample” to try example data first.
Adjust options and run：Adjust the options if needed, then click the process / convert / generate button.
Get the result：Everything runs locally in your browser; copy or download the result — your files are never uploaded.

FAQ

What is the difference between live mode and file mode?: Live mode streams: it transcribes while capturing the microphone, using either the local WASM engine (works offline) or the browser cloud service (needs network, audio is sent to the browser vendor). File mode transcribes a local audio/video file — fully local, no microphone needed.
Where do the who / when / what fields come from?: When the model outputs JSON (Start/End/Speaker/Content), speaker and timestamps come straight from the model; with plain text, sentences are split and the segment range is distributed by character count, with the speaker falling back to "Speaker 1". The data layer fills any gap: a missing end time becomes start + 2 s, empty content becomes "(no content)", so every block always has all three fields in the fixed order who → when → what.
How does live streaming produce text while I speak?: It uses chunked streaming: once the buffer reaches the window length (5 s by default) or a pause is detected, that slice is sent to the engine and the result is de-duplicated against the previous chunk before being appended. It is not token-level streaming, but text appears at every pause — latency equals the window size.
What is 3A and why do some switches not take effect?: 3A means AEC (echo cancellation), ANS (noise suppression) and AGC (auto gain), provided by the browser's built-in WebRTC audio module. This tool toggles them via getUserMedia constraints and reads the result back with getSettings(). Whether they apply depends on browser and device — the panel shows the effective values, and unsupported extras (voice isolation, AEC mode) are skipped automatically.
Why is the first transcription slow?: The first run downloads two GGUF models (VAE encoder ~0.65 GB and LM decoder ~0.92 GB, ~1.6 GB total) with live progress; they are then cached in the browser, after which the panel shows "cached (works offline)" and no network is needed.
What engine does offline recognition use?: Microsoft's VibeASR.cpp (built on llama.cpp) is compiled to WASM with Emscripten. The wasm/ directory here holds the wrapper; run npm run build:wasm to build it yourself. Inference runs in a Web Worker so the page stays responsive.
Any caveats running a 1.6 GB model in a browser?: Weights live in WASM memory: use a desktop browser with 8 GB+ RAM and keep the tab in the foreground. First load takes tens of seconds; afterwards each transcription only processes the audio.
What if the transcript is inaccurate?: Use a quiet room and stay 20–40 cm from the mic; enable ANS for background noise and AGC for level; split long recordings; accuracy improves when the selected language matches what is spoken.

More tools in Local AI

Local Text-to-Speech — Synthesize Chinese and English text entirely in the browser with the Hojo-TTS family neural models (Hojo-TTS-Light and Hojo-TTS-Light-40M, switchable from a dropdown): no network, audio never leaves the device. Weights (~250MB, depending on the chosen model) download once, then it works offline; multiple built-in voices (Chinese + English).

返回崽崽工具箱首页