Superwhisper 为私密的设备端听写添加了 Cohere Transcribe

Why it matters

The integration gives Cohere an end-user channel for its open speech model and lets Superwhisper improve private voice input without training its own frontier ASR system.

Neil Chudleigh (@neilsuperduper), the founder of Superwhisper, has added Cohere's open-weight Transcribe speech recognition model to the app for local, offline dictation. Superwhisper announced the integration on July 28th and promoted the partnership in a post on X on July 29th. (superwhisper.com)

Superwhisper 的创始人 Neil Chudleigh (@neilsuperduper) 已将 Cohere 的开源权重语音识别模型 Transcribe 添加到该应用中，用于本地离线听写。Superwhisper 于 7 月 28 日宣布了该集成，并在 7 月 29 日通过一则 X 帖子 宣传了这次合作。(superwhisper.com)

The addition advances the product thesis Chudleigh has pursued since building Superwhisper for himself: voice input should work across applications without sending every recording to a remote server. Chudleigh previously co-founded PartnerStack, the B2B partner-management platform that went through Y Combinator in 2015 and was acquired by AppDirect on April 14th, 2026. (appdirect.com)

此举推进了 Chudleigh 自为自己构建 Superwhisper 以来一直秉持的产品论点：语音输入应能在各类应用中工作，而无需将每条录音发送到远程服务器。Chudleigh 之前共同创立了 PartnerStack —— 一个 B2B 合作伙伴管理平台，该公司于 2015 年参加了 Y Combinator，并于 2026 年 4 月 14 日被 AppDirect 收购。(appdirect.com)

Chudleigh has said he started Superwhisper after finding Apple's dictation inaccurate and inconsistent across apps. He initially built the product around offline models, accepting the added engineering work required to manage large models and audio processing on consumer hardware. That architecture also gave Chudleigh a clear position for sensitive legal, medical, government and software-development workflows where users may be unwilling or unable to upload speech. (deepgram.com)

Chudleigh 表示，他在发现 Apple 的听写在不同应用间不准确且不一致后启动了 Superwhisper。他最初将产品构建在离线模型之上，接受了在消费级硬件上管理大型模型和音频处理所需的额外工程工作。这种架构也为 Chudleigh 在法律、医疗、政府和软件开发等敏感工作流中提供了明确的立足点——这些场景下用户可能不愿或无法上传语音。(deepgram.com)

A 1.3GB model inside the app

应用内的 1.3GB 模型

Users can install Cohere Transcribe from Superwhisper's model library as a 1.3GB download, select it as their voice model and dictate without an internet connection. Superwhisper says the audio remains on the user's machine. The integration also applies Superwhisper's vocabulary layer, which is designed to recognize user-defined names, product terms, snippets and technical jargon. (superwhisper.com)

用户可以从 Superwhisper 的模型库中安装 Cohere Transcribe，作为一个 1.3GB 的下载，选择它作为语音模型并在无网络连接下进行听写。Superwhisper 表示音频会保留在用户的设备上。该集成还应用了 Superwhisper 的词汇层，旨在识别用户自定义的名称、产品术语、代码片段和专业术语。(superwhisper.com)

Cohere released Transcribe on March 26th as a 2 billion-parameter Conformer model trained for 14 languages, including English, French, German, Spanish, Mandarin, Japanese, Korean and Arabic. The model weights are available under the Apache 2.0 license, giving developers permission to use and modify the model under the license's terms. (cohere.com)

Cohere 于 3 月 26 日发布了 Transcribe，这是一个拥有 20 亿参数的 Conformer 模型，针对包括英语、法语、德语、西班牙语、普通话、日语、韩语和阿拉伯语在内的 14 种语言进行训练。该 模型权重 在 Apache 2.0 许可证下可用，允许开发者在该许可证条款下使用和修改模型。(cohere.com)

Cohere's benchmark table placed Transcribe first on the Hugging Face Open ASR Leaderboard at its March launch, with an average English word error rate of 5.42%. The table showed OpenAI's Whisper Large v3 at 7.44% and ElevenLabs Scribe v2 at 5.83%. Those figures describe a standardized collection of English datasets rather than every dictation environment, accent or microphone setup. (cohere.com)

Cohere 在 3 月发布时的基准表将 Transcribe 排在 Hugging Face Open ASR 排行榜首位，平均英文词错误率为 5.42%。表中显示 OpenAI 的 Whisper Large v3 为 7.44%，ElevenLabs 的 Scribe v2 为 5.83%。这些数字描述的是一组标准化的英文数据集，而非涵盖所有听写环境、口音或麦克风设置。(cohere.com)

The model also carries practical constraints. Cohere's model card says Transcribe performs best when users select one supported language in advance. It lacks automatic language detection, timestamps and speaker diarization, and Cohere warns of inconsistent results on code-switched audio. The model can also attempt to transcribe background noise unless an application places voice-activity detection or a noise gate in front of it. Superwhisper's vocabulary wrapper addresses specialized words, though it does not remove those underlying model limitations. (huggingface.co)

该模型也存在实际限制。Cohere 的模型卡指出，Transcribe 在用户预先选择一种受支持语言时表现最佳。它不具备自动语言检测、时间戳和说话人分离功能，Cohere 警告在语言切换（code-switched）的音频上结果可能不稳定。除非应用在模型前端加入语音活动检测或噪声门，否则模型也可能试图转录背景噪音。Superwhisper 的词汇包装器可处理专业词汇，但不能消除这些底层模型的局限性。(huggingface.co)

Cohere gets an end-user distribution channel

Cohere 获得了面向终端用户的分发渠道

For Superwhisper, the partnership adds a competitive open model without requiring Chudleigh to train a frontier speech system from scratch. Superwhisper can instead compete through model selection, local inference, vocabulary controls and the workflow that inserts finished text into email, coding and productivity applications.

对 Superwhisper 来说，这次合作增加了一个具有竞争力的开源模型，而无需 Chudleigh 从零开始训练一套前沿语音系统。Superwhisper 可以通过模型选择、本地推理、词汇控制以及将生成文本插入电子邮件、编码和生产力应用的工作流来进行竞争。

For Cohere, Superwhisper puts Transcribe directly in front of people using speech as a daily computer input, four months after the model's release. Cohere introduced Transcribe as the first step toward speech intelligence inside North, its enterprise agent platform. Distribution through a consumer-facing dictation app gives the model a separate route into real workflows while preserving Cohere's emphasis on private deployment and infrastructure control. (cohere.com)

对 Cohere 而言，在模型发布四个月后，Superwhisper 将 Transcribe 直接呈现在那些将语音作为日常电脑输入的人面前。Cohere 将 Transcribe 作为其企业代理平台 North 中实现语音智能的第一步。通过面向消费者的听写应用进行分发，为该模型进入真实工作流提供了一条独立途径，同时保留了 Cohere 对私有部署和基础设施控制的重视。(cohere.com)

The Toronto connection also gives both companies a sovereignty pitch. Superwhisper describes the integration as a Canadian model running inside a Canadian-built application, with processing kept on the user's hardware. That framing is useful for organizations concerned about where audio is processed, though the immediate product benefit is simpler: users can install a comparatively accurate speech model without configuring Python, downloading weights manually or operating inference infrastructure.

多伦多的关联也为双方提供了一个主权诉求点。Superwhisper 将这次集成描述为一个在加拿大构建的应用内运行的加拿大模型，并且处理保留在用户的硬件上。对于关心音频处理地点的组织来说，这种表述很有用，尽管即时的产品好处更简单：用户可以安装一个相对准确的语音模型，而无需配置 Python、手动下载权重或运行推理基础设施。

Chudleigh's bet is that voice becomes a general input layer for software rather than a separate transcription destination. Adding Cohere Transcribe strengthens the part of that argument users encounter first. Dictation only saves time when the transcript requires fewer corrections than typing the sentence would have taken.

Chudleigh 的押注是，语音将成为软件的通用输入层，而不是一个独立的转录终点。添加 Cohere Transcribe 强化了用户最先接触到的那部分论点。只有当听写生成的文本所需的修正少于键入该句所需的时间时，听写才真正省时。