dsh-plugin-rapidspeech

dsh-plugin-rapidspeech

English | [简体中文](README.zh.md)

Local, offline speech-to-text for the [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) (dsh). Click the microphone in the composer, speak, and the recognized text is inserted into your draft — transcribed by a [RapidSpeech.cpp](https://github.com/RapidAI/RapidSpeech.cpp) rs-server running on your own machine. Audio never leaves your computer.

Features

• Fully offline — audio, model inference, and transcripts stay on this machine end to end.
• Managed server — on first use, dsh downloads the rs-server binary and your chosen model, verifies their integrity, spawns the server on a loopback port, and reclaims it after idle time. No manual services.
• Model switching from the plugin page — pick a model in the plugin's settings; it downloads on the next recording and hot-swaps. Every downloaded model stays cached, so switching back is instant.
• Curated model catalog — X-ASR, SenseVoice, and FunASR-Nano from RapidAI's official mirrors (ModelScope, HuggingFace, hf-mirror), each pinned to a SHA-256.
• Coexists with other providers — installs beside the official SenseVoice plugin; pick a provider per recording from the voice-input selector.
• External mode — point the plugin at an rs-server you run yourself; the plugin then only connects and never manages it.

Requirements

• DeepSeek Harness 0.2.2 or newer (npm CLI or Desktop app)
• CPU with AVX2 — Intel Haswell (2013) / AMD Excavator (2015) or newer. Both the

Linux and Windows builds target AVX2 as their baseline, so the requirement is

identical on either OS.

• Disk: ~200 MB–1 GB per model; RAM: ~1–2 GB while transcribing
• Windows x64 and Linux x64

Install

From the plugin marketplace (Desktop / Web UI)

Two ways — both put RapidSpeech ASR (local) in your plugin list:

• Search and install (simplest): open the market's install page, search for

dsh-plugin-rapidspeech, and install it. The package lives on the official npm

registry; if your npm mirror lags behind, give it a few minutes or point the

registry back at registry.npmjs.org.

• Index source (curated card with a description): add this index URL in the

market's index settings, then find RapidSpeech ASR (local) and install:

  https://github.com/yoletPig/dsh-plugin-rapidspeech/releases/latest/download/market-index.json

From the CLI

dsh plugin --profile  add dsh-plugin-rapidspeech

Or install a specific release artifact without any build tooling:

# Download dsh-plugin-rapidspeech-.tgz from GitHub Releases, then:
dsh plugin --profile  add ./dsh-plugin-rapidspeech-.tgz

Installing from a git checkout

git clone https://github.com/yoletPig/dsh-plugin-rapidspeech.git
cd dsh-plugin-rapidspeech
npm install          # the prepare script builds lib/
dsh plugin --profile  add ./dsh-plugin-rapidspeech

On pnpm ≥ 10, the first add from git asks you to allow the build script — add the

package key printed by pnpm to the profile's pnpm-workspace.yaml under allowBuilds.

Quick start

1. Start dsh and open the composer.

2. Click the microphone, then Install when prompted — the rs-server binary and the

model download in the background (one-time, ~180 MB for the default model).

3. Record. Stop, and the transcript appears in your draft, ready to edit and send.

The server reclaims itself after 5 idle minutes and wakes in about 2 seconds on the

next recording.

Models

| modelId | Size | Languages | Mirrors |

| --- | --- | --- | --- |

| xasr-q8_0 (default) | 178 MB | auto, zh, en | ModelScope |

| sense-voice-small-q4_k | 182 MB | auto, zh, en, yue, ja, ko | HuggingFace, hf-mirror, ModelScope |

| funasr-nano-q4_k_m | 692 MB | auto, zh, en | HuggingFace, ModelScope |

Switching models

Open the plugin's page in dsh and pick a model — the field is live: the next

recording downloads it (after your consent) and subsequent recordings use it.

All catalog models live in the same cache, so switching back to a previously

downloaded model needs no download.

X-ASR is only published on ModelScope, so the HuggingFace mirror is not offered

for it; choosing an unavailable mirror fails explicitly rather than falling back.