← Audio

WhisperWhisper

OpenAI（开源）OpenAI (open source)

开源MITOpen source, MIT 本地转写Local transcription 多语言Multilingual 模型非产品Model, not a product 隐私向Privacy-focused partial lifecycle: active 核验Checked 2026-10-22

Whisper 是转写领域的「开源基座」：MIT 权重多档位本地跑，隐私敏感音频不出本机；它是模型不是产品——会议记录体验要壳件，说话人分离靠衍生生态。

Whisper is the open-source base of transcription: MIT weights in multiple sizes run locally, privacy-sensitive audio never leaves the machine; it's a model, not a product — meeting-note UX needs a shell, diarization needs derivative tooling.

适合与不适合When to use it, when not to

适合：隐私敏感转写、离线环境、开发者集成、多语言。不适合：开箱会议产品、实时字幕开箱、无 GPU 重度使用。

Fits: privacy-sensitive transcription (audio stays local), offline environments, developer integration (MIT), multilingual. Not a fit: out-of-box meeting-product UX (it's a model), out-of-box diarization, heavy use without a GPU.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-22

出品属地Origin: 国外出品foreign-made　OpenAI（美国）开源项目。OpenAI (US) open-source project.
国内使用Direct access: 大陆可直连direct OK　github.com/openai/whisper 大陆直连实测 200（2026-10-22）；权重经 Hugging Face 分发，HF 大陆可达性未在本档案实测。github.com/openai/whisper measured 200 direct from mainland (2026-10-22); weights are distributed via Hugging Face, whose mainland reachability was not measured in this profile.
备案状态CAC filing: 未见于备案名单not in filing list：开源模型不适用服务备案（不面向公众提供服务）；OpenAI 未见于网信办公示名单（2026-10-22 检索）。an open-source model is not subject to service filing (it serves no public service); OpenAI is not in the CAC public registry (searched 2026-10-22).

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 开源转写模型族（tiny→large-v3 多档，官方 README 列模型尺寸与语言覆盖）；权重经 Hugging Face 分发（官方 README 链接）。衍生生态庞大：whisper.cpp、faster-whisper、WhisperX 等（官方 README 提及部分——生态全貌未逐条清点）。
An open-source transcription model family (tiny→large-v3 multiple tiers, the official README lists model sizes and language coverage); weights distributed via Hugging Face (linked from the official README). A huge derivative ecosystem: whisper.cpp, faster-whisper, WhisperX etc. (partially mentioned in the official README — the ecosystem's full shape not inventoried item by item).
运行位置Where it runs: 本地运行（Python + PyTorch，官方 README 安装面）；GPU/CPU 皆可，档位换速度/质量。托管线：OpenAI API 的 whisper 转写端点（platform.openai.com 文档线）是同一模型的云形态。
Runs locally (Python + PyTorch, the official README's install surface); GPU or CPU both work, tiers trade speed/quality. Hosted line: the whisper transcription endpoint on the OpenAI API (platform.openai.com docs line) is the cloud form of the same model.
本地文件Local files: 完全本地：音频、转写结果都在本机——这是它与所有云端转写产品最根本的数据边界差异。
Fully local: audio and transcription results stay on your machine — this is its most fundamental data-boundary difference from every cloud transcription product.
关机后的任务Tasks after shutdown: 批处理任务制（文件进、文本出，本机跑完即退）；无后台常驻/记忆概念。
Batch task-based (file in, text out, runs and exits locally); no background-resident/memory concept.
工具与扩展Tools & extensions: CLI 与 Python API（官方仓）；时间戳/翻译（译到英文）能力内置（官方 README）； VAD/说话人分离不在官方仓——靠 WhisperX 等衍生件（衍生面未逐条核验）。
CLI and Python API (official repo); timestamps/translation (into English) built in (official README); VAD/speaker diarization are not in the official repo — they rely on derivatives like WhisperX (derivative surface not verified item by item).
上下文与记忆Context & memory: 无上下文概念——逐段转写；长音频靠分段策略（官方仓有长音频处理说明）。
No context concept — segment-by-segment transcription; long audio relies on segmentation strategies (the official repo has long-audio handling notes).
权限与限制Permissions & limits: 本地跑无账号无上传——隐私敏感音频（医疗/法务/私人）走本地是它的不可替代场景；云端 API 形态则按 OpenAI 平台条款（未逐条核验）。
Running locally means no account, no upload — privacy-sensitive audio (medical/legal/personal) going local is its irreplaceable scenario; the cloud API form follows OpenAI platform terms (not verified item by item).
适合什么任务Best for: 适合：隐私敏感的转写（音频不出本机）、离线环境、开发者集成（MIT 无许可负担）、多语言转写。不适合：开箱即用会议记录产品体验（它是模型不是产品）、说话人分离开箱（要衍生件）、无 GPU 重度使用（大档位 CPU 慢）。
Fits: privacy-sensitive transcription (audio never leaves your machine), offline environments, developer integration (MIT, no licensing burden), multilingual transcription. Does not fit: out-of-the-box meeting-notes product experience (it is a model, not a product), out-of-the-box speaker diarization (needs derivatives), heavy use without a GPU (big tiers are slow on CPU).

头号误解Top misconception

Whisper 是模型不是产品——「会议转写工具」体验（Otter 类）要自己搭壳或用衍生产品Whisper is a model, not a product — 'meeting transcription tool' UX (Otter-type) means building your own shell or using a derivative product
说话人分离/实时流不在官方仓——看到「Whisper 实时字幕」类产品都是衍生件，质量与许可各自核Diarization/live streaming are not in the official repo — 'Whisper live captions' products are all derivatives; verify quality and license individually
官方仓与 OpenAI API 的 whisper 端点是「同模型两形态」——本地 MIT vs 平台条款，数据边界完全不同The official repo and the OpenAI API whisper endpoint are one model in two forms — local MIT vs platform terms, completely different data boundaries

未知项清单Known unknowns

衍生生态全貌（whisper.cpp/faster-whisper/WhisperX 等未逐条核验质量与许可）
OpenAI API 端点当前模型版本与本地仓版本是否同步（未核验）
官方后续模型线走向（未核验到声明）

The derivative ecosystem's full shape (whisper.cpp/faster-whisper/WhisperX etc., quality and licenses not verified item by item)
Whether the OpenAI API endpoint's current model version matches the local repo version (not verified)
The official successor model line's direction (no statement verified)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-22，内容更新日content updated on 2026-10-22。

类型Type	名称Name	链接Link
official	Whisper · 官方仓（README/模型清单/LICENSE）Whisper · official repo (README/model list/LICENSE)	https://github.com/openai/whisper
changelog	Whisper · 官方仓 Releases（changelog）Whisper · official repo Releases (changelog)	https://github.com/openai/whisper/releases
official	OpenAI · Whisper 发布页（模型叙事）OpenAI · Whisper launch page (model narrative)	https://openai.com/index/whisper/

实测记录Measured runs

本档未跑音频质量与延迟测试；结论只覆盖能力面与许可面。

实测建议（三条，都指向官方可读页没给的东西）：

中文转写在中/英混说场景的档位选择（large-v3 vs 衍生微调件）
whisper.cpp 在消费级 GPU 的实时率
与 Deepgram/AssemblyAI 云端 API 的中文准确率对比

This site has not completed measured runs. The coding-oriented test protocol is at protocols/coding/_protocol.md (the coding task suite does not apply to creation products).

Suggested measurements (three, all pointing at things the official readable pages do not give):

Tier choice for Chinese transcription in Chinese/English mixed scenes (large-v3 vs fine-tuned derivatives)
whisper.cpp's real-time factor on consumer GPUs
Chinese accuracy vs the Deepgram/AssemblyAI cloud APIs