@huliux/dsh-asr-plugin
适用于 macOS（仅限 Apple Silicon）的 DSH 本地录音和会议转录
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:huliux/dsh-asr-plugin#1c03c3038e29e565c315c8e10d10e12b8586ec8a
Asr Plugin 以 @huliux/dsh-asr-plugin 发布，当前版本为 0.1.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 简体中文
Unofficial local meeting transcription plugin for DeepSeek Harness (DSH) on Apple Silicon macOS. The plugin imports WAV, M4A and MP3 files, records microphone and system audio, and produces transcripts with timestamps and speaker labels. Meetings can be referenced explicitly in DSH messages for reading and export.
| Component | Requirement |
|---|---|
| Host | DSH 0.2.0-rc.2 |
| Runtime | Node.js 24 |
| Platform | macOS, Apple Silicon (darwin / arm64) |
| Microphone API | macOS 13.5 or later |
| System audio and dual-track APIs | macOS 14.2 or later |
These macOS versions specify API availability; they are not a tested OS matrix. DSH desktop and Web modes both process audio on the Mac running DSH. Windows, Linux and Intel Macs are unsupported.
Install @huliux/dsh-asr-plugin from the DSH Plugins page. Web profiles also
support the CLI:
dsh plugin --profile web add --ignore-scripts @huliux/dsh-asr-plugin@0.1.2
The package includes native add-ons and an ad-hoc signed recording Helper. The Helper is not notarized; macOS may require permission or security approval at installation and after updates. Keep system security protections enabled.
Desktop and Web profiles share the default meeting store. Run one plugin host per data directory. Preserve meeting and model data when replacing a package.
For package verification, local archive installation and release channels, see distribution. Source downloads require the complete source build before recording. Report installation problems through GitHub Issues, including the plugin, DSH, Node.js and macOS versions without meeting data or credentials.
In plugin settings, download the base models (approximately 278 MiB). Optional punctuation models require approximately 274 MiB and apply automatically to new recordings, imports and retranscriptions after verified installation. Existing tasks and transcripts retain their processing mode. A damaged installed punctuation pack must be repaired before starting new tasks.
Models are downloaded from pinned upstream revisions and verified by size and SHA-256 before installation. Weights are excluded from Git, npm and GitHub Releases. Enabling the plugin does not initiate a model download. Downloads support cancellation and retry; complete verified files can be reused after failure. See model assets for sources, proxy behavior and offline staging.
Use Recording permissions in plugin settings to check microphone and system audio access. Each recording start performs fresh admission checks. A failed system-audio check may reflect permission, output volume or device routing. Silence during a recording does not establish a permission denial.
Expanding the recorder prepares models without capturing audio. Sequential recordings can reuse loaded model resources, which are released after five idle minutes. Audio, drafts and speaker state remain separate for each meeting.
Select a DSH session to record, or ask the DSH agent to import an absolute local
audio path. Use the @ picker to reference a meeting and request reading or export.
File writes use DSH Approval. Meetings persist independently of DSH sessions and
workspaces.
Audio inference runs locally after model preparation. Reading a transcript with an LLM can send its content to the provider configured in DSH; host configuration and provider terms govern that operation.
See development for source builds and verification, CONTRIBUTING.md for contributions, and publishing for package preparation.
Project-owned source is licensed under Apache-2.0. Third-party source
and models retain their respective terms; see
THIRD_PARTY_NOTICES.md and the notices in third_party/
and vendored native directories.