dsh-read-aloud

One line: text-to-speech for DeepSeek Harness. Every Assistant reply gets a 🔊 button that reads it aloud in a MiniMax voice. Optional auto-read each turn, and it stops the moment you send a new message.

Turns DSH replies into audio with MiniMax synthesis — pick any system voice, or reuse a cloned one.

Who it is for: long replies are tiring to read, you want to listen while doing something else, you are on a phone, or you need the answer read out to someone next to you.

What you get	Detail
🔊 one button per reply	In the same strip as copy / fork / 👍👎. Click to read, click to stop, click again to replay (replays are free)
Auto-read per turn	Optionally read the last paragraph of every reply
Barge-in	Sending a new message cuts the audio immediately (Doubao-style); the next reply is read as usual
Voice and rate	MiniMax system or cloned voices, model, speed and emotion, with a preview
Code stays silent	Only prose text blocks are read — no code, tool calls or reasoning

Synthesis goes through MiniMax's cloud API and needs network access; the DSH UI itself stays local.

A community plugin, not an official DeepSeek or MiniMax product.

English | 简体中文

---

The gap it fills

DeepSeek Harness ships no speech synthesis at all. The community plugin dsh-minimax-voice brings MiniMax cloning, TTS and ASR — but its playback lives in agent tools and a settings panel, with no "click a reply and hear it".

This plugin is the thin layer that connects a message action button to existing synthesis. With dsh-minimax-voice installed it reuses that plugin's key, voice library and billing; without it, the built-in standalone mode talks to MiniMax directly — one pasted API key and it runs.

Features

• A speaker button per reply, in the same strip as copy / fork / 👍👎: click to read that message, click while it plays to stop, click again to replay.
• One reading per turn. A multi-step turn emits one Assistant message per step; reading each would talk over the tool calls, so only the newest prose is kept and spoken when the turn closes.
• Optional auto-read: speak the end of every turn.
• Barge-in: sending a new message cuts the current audio immediately (Doubao-style); the next reply is read as usual. Toggle in settings.
• Code and links are skipped: fenced blocks are dropped whole, links contribute only their label, and images or table rules never get read. Inline code such as npm_install is kept — that is usually the name you need to hear.
• History works too: the event stream is cached silently on page load, without speaking it.
• Single playback: starting another message stops the current one.
• Configurable voice: any MiniMax system voice or a cloned voice_id, plus model and speed, with a preview button.

Requirements

Either of:

• Recommended: install dsh-minimax-voice. This plugin's host calls its minimaxVoice service directly (same process, no HTTP); the key, cloned voices and billing stay there.

• Standalone mode: the host calls MiniMax itself. Key priority:

  1. config.apiKey in the plugin configuration
  2. the MINIMAX_API_KEY environment variable (restart the Harness after setting it)
  3. saved from the settings panel (plain text at ~/.dsh/read-aloud/secret.json)

The browser only ever calls this plugin's own routes:

GET  /read-aloud/api/status                                  ->  { configured, backend, keySource }
POST /read-aloud/api/tts    { text, voiceId, model, speed }  ->  { path, file, durationMs, characters, backend }
POST /read-aloud/api/key    { key }                          ->  { configured }
GET  /read-aloud/api/output/:file                            ->  mp3

Why not fetch dsh-minimax-voice's HTTP surface from the browser? That route runs every request through the DSH connection trust check, which answers 401 to a plain fetch. The client then decided "no backend" and failed silently — the bug 0.2.1 fixes. Calling the service in-process removes auth, CORS and ports from the equation.

With neither backend configured, the button shows ⚠️ and puts the concrete error in its tooltip instead of failing silently.

Install

Source install only for now — the package is not on npm yet, so the dsh plugin add npm form will not work.

git clone https://gitee.com/op00007/dsh-minimax-autoread.git dsh-read-aloud
dsh plugin --profile desktop add ./dsh-read-aloud

Already have a local copy:

dsh plugin --profile desktop add /absolute/path/to/dsh-read-aloud

Refresh the page afterwards — client plugins mount at page load.

Usage

1. Hover any Assistant reply; a 🔊 appears in its action strip.
2. Click to read it. While playing the icon becomes ⏹; click to stop, click again to replay.
3. Settings → MiniMax Auto-read holds the auto-read switch, the voice id, the model, the speed and a preview button.

Settings

Stored in the browser's localStorage under dsh.read-aloud.settings. Nothing is polled, reported, or written into host configuration.

Key	Default	Meaning
enabled	true	Speak the end of each turn
interrupt	true	Cut playback when a new message is sent
skipMarkup	true	Drop code blocks and link addresses
voiceId	female-shaonv	MiniMax voice id, system or cloned
model	speech-2.8-hd	Synthesis model (2.8 / 2.6 / 02 / 01, hd and turbo)
speed	1	Speech rate, 0.8–1.5

How it works

   ┌──────────────────────────── browser ───────────────────────────┐
   │                                                                │
   │  session event stream: binding.eventSource                      │
   │        │                                                       │
   │        ├── assistant/message ──▶ texts.set(message.id, prose)   │
   │        │                                                       │
   │        └── turn/end ──▶ auto-read on? ──▶ speak(pending)        │
   │                                                                │
   │  slot conversation.chat.assistant-actions ──▶ [🔊 button]       │
   │        └── click ──▶ toggle(messageId)                          │
   │                        │                                       │
   │                        ▼                                       │
   │              POST /mmvoice/api/tts ──▶ { path }                 │
   │                        │                                       │
   │                        ▼                                       │
   │              new Audio('/mmvoice/api/output/<file>')            │
   └────────────────────────────────────────────────────────────────┘

Design notes:

• The host resolves the backend; the browser never authenticates. The client only calls /read-aloud/api/*; the host prefers the minimaxVoice service and falls back to MiniMax directly. Produced mp3s are copied into this plugin's output directory so one stable URL serves them.
• Prose is cached by messageId. The action slot hands over an id only, so the button has to learn what that message said. The event stream is the only reliable source.
• Replay is not speech. Page load replays history; the first pump runs with synced still false, so it fills the cache without talking.
• Replay is free. Synthesized audio is cached in memory per message id; a second click plays the file instead of re-billing.
• No DOM fallback. When the prose is missing the button reports it. An early version scraped the nearest [data-message-bubble] element — that attribute marks the user bubble, so the fallback would have read your own message back to you.
• Only text blocks are spoken. Reasoning, code and tool calls stay silent.

Compatibility

• Developed and verified against DSH 0.2.2-rc.1 (desktop).
• engines.dsh is >=0.2.2-rc.1 <0.3.0-0.
• Depends only on shared modules the web shell provides: react and @deepseek-ai/dsh-client-ui-slots.
• Uses the slots conversation.chat.assistant-actions and settings.section. If a later release renames them, the button silently disappears and this plugin needs updating.

Known limitations

• Standalone voice suggestions are hardcoded common system voices; cloned voices are only available through dsh-minimax-voice.
• Prose comes from the event stream. A message whose events are not in the current stream (older paged-out history, for example) will not be read; the tooltip says the text is not cached yet.
• Browser autoplay policy. The first playback needs a prior user interaction on the page; if blocked, the failure shows in the tooltip.
• Single playback, no queue. Replies are not chained.
• Truncated at 8000 characters (MiniMax accepts 10000 per call).
• The replay cache is memory only. After a page reload, replaying the same message synthesizes again and bills again.

Development

dsh-read-aloud/
├── client.js            # all logic: slots, event subscription, synthesis, playback
├── index.js             # host placeholder, declares the plugin
├── cordis.patch.yml     # inserts the read-aloud row
├── package.json
├── icon.svg
├── README.md
├── README.zh-CN.md
├── CHANGELOG.md
└── LICENSE

A file: install is a copy, not a symlink: editing the checkout does not update an installed profile. Reinstall, or overwrite node_modules/dsh-read-aloud/ and refresh.

License

MIT