DeepSeek Harness Plugin

OldQin-Family/dsh-breeze-tts

Stars ★ 1 Category Voice & Audio Added 2026-11-20

DSH client for a self-hosted Breeze TTS service: AI auto-read plus a voice workshop for designing, cloning, and managing voices. Requires a running Breeze service; no model is bundled.

Install

# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)

dsh plugin --profile web add github:OldQin-Family/dsh-breeze-tts

Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).

README

English · 中文

dsh-breeze-tts · Breeze Voice

Version 1.3.0 · a DeepSeek Harness (DSH) plugin: make AI speak and create the voice you want.

This plugin does not contain a speech model. It is a client: you run your own Breeze TTS service, and the plugin talks to it. No service, no sound.


60-second overview

Once installed, the DSH Settings menu shows two entries for this plugin:

Entry What it does In short
Breeze Settings (or Local voice if a legacy TTS plugin is also enabled) Auto-read AI replies, pick a voice, adjust speed Use a voice
Breeze 音色工坊 (Breeze Voice Workshop) Design a voice from text / clone from a recording / manage voices Make a voice

Create a voice first, then let it speak — voices designed in the Workshop are available in Breeze Settings.


How it works

Three layers: your DSH → this plugin (client) → your Breeze TTS service.


Interface

Three real interface captures from a local run, unedited; no generated images are bundled.


Contents


Step 0 — Check your hardware first

The plugin itself barely uses hardware. All hardware requirements are on the Breeze TTS service side.

Your machine Can it run Which build
NVIDIA GPU, ≥ 4 GB VRAM ✅ Comfortably C++ build, Q8_0
AMD / Intel GPU (incl. iGPU) ✅ Yes C++ build (Vulkan — no CUDA/ROCm needed)
CPU-only system ⚠️ Works, but slower than real time C++ build, Q4_K
Pre-2013 x64 CPU ❌ No No AVX2, no fallback
Apple Silicon / ARM ❌ No No prebuilt package

Bottom line: most users should choose the C++ build. It supports NVIDIA, AMD and Intel GPUs via Vulkan, as well as CPU-only systems (which run slower than real time). A reasonably recent graphics driver is needed for Vulkan. 4 GB VRAM is enough for Q8_0; Q4_K can fit in about 3 GB. The one hard requirement is AVX2 on the CPU.

Full data and the 5-step AI decision flow: HARDWARE.md.


Step 1 — Run a Breeze TTS service

The plugin is a client — without a service it makes no sound.

  • Recommended: Breeze-TTS-2.cpp (Windows/Linux, NVIDIA/AMD/Intel, CPU fallback)
  • Model: Hugging Face — download the audio_tokenizer/ directory too, or there will be no sound
  • Q8_0 (3.3 GB) if you have ≥ 4 GB VRAM; Q4_K (2.4 GB) around 3 GB
  • Start: breeze-server breeze-tts-2-q8_0.gguf --webui (default port 8080)
  • Verify: http://127.0.0.1:8080/health returns {"status":"ok",...}

The official PyTorch build also works, but it has no /v1/voices (voice dropdown will be empty; you can still type an id), and it needs Linux + NVIDIA CUDA with 12 GB+ VRAM. C++ is recommended.

⚠️ The Breeze service itself provides no authentication or rate limiting — do not expose it to the public internet.


Step 2 — Install and configure the plugin

Easiest: ask your AI assistant. Copy the block that matches your situation and paste it into Claude, Codex, Cursor, or another assistant. It will read README-AI.md and guide you through setup, showing you every command and response.

Case A — Installed in DSH, not configured yet

The DSH plugin dsh-breeze-tts is installed but not configured. Please read
README-AI.md in the plugin directory, then set it up:
1. GET /dsh-breeze-tts-api/health — paste the raw response;
2. My Breeze service is at http://<your-IP>:8080 — PUT /config to write it
   into endpoint, then GET /config to read it back;
3. GET /voices — paste the list, I'll tell you which one to set;
4. POST /speak-preview with "Hello, this is a test" — prove it actually speaks.
Paste every command and raw response, don't just say "done".

Case B — Just downloaded, not installed yet

I downloaded a DSH plugin dsh-breeze-tts to <path>, not installed yet.
Please read README-AI.md and install it into DSH:
1. Check whether @dsh-external/dsh-plugin-local-ai-tts and dsh-breeze-voices
   are installed — if so, tell me first, don't uninstall by yourself;
2. Move the directory to a pure-ASCII path (Chinese paths break pnpm);
3. Follow README-AI.md §1 to edit the profile's package.json and run pnpm install;
4. Verify node_modules files match the source fingerprints — stop if not;
5. I'll restart DSH, then continue with Case A.
Paste every command and output, don't skip steps.

Case C — Not sure the machine can run it

Please read HARDWARE.md and judge whether this machine can run Breeze TTS:
1. Measure architecture, GPU vendor/model, actual VRAM, CPU, Vulkan support;
2. Follow HARDWARE.md §5 and tell me: which build, which quant level,
   and roughly how fast (or honestly "slower than real time");
3. Paste every command and raw response — don't just say "it's fine";
4. Don't invent numbers the official docs don't give — say "not published".

Step 3 — Use it

DSH → Settings → Plugins → Breeze Settings (or Local voice in coexistence mode) / Breeze 音色工坊

  • To hear replies: open Breeze Settings (or Local voice if a legacy TTS plugin is also enabled), turn on Auto Read, and pick a voice. If nothing plays, click anywhere on the page first, then preview (browsers block autoplay).
  • To make a voice: open Breeze 音色工坊 (Breeze Voice Workshop).
    • Design: write a description (e.g. "a soft-spoken young woman in her early twenties, slightly breathy, smiling as she talks") → generate → preview → name and save.
    • Clone: upload a recording + type exactly what was said → save.
    • Voice names may only use [A-Za-z0-9_-]; Chinese names are rejected by Breeze (HTTP 400).

Writing a good description:

  1. One sentence, not a parameter list.
  2. Cover "who + what the voice sounds like + how they speak" (age, gender, texture, mood, pace).
  3. Write the description in the language being spoken — official requirement.
  4. Want it to follow the description more closely? Raise cfg_scale (default 1.0, try 2–4; above 3 can sound harsh).

Event tags: use parentheses in English — (laugh), (sigh) — and square brackets in Chinese — [笑], [叹气]. They usually need cfg_scale 2–3 to trigger reliably.

Hear official examples at breezeblue.ai (home page, seven categories — there is no separate voice library page).


Cross-machine access

By default, only local requests are allowed. To let a DSH instance call a Breeze TTS service running on another machine:

  • remoteAccess: when off, non-local requests get 403;
  • trustedLocalCidrs: addresses here are treated as local; when requireLocalToken=false, they can read and write without a token. Adding an IP means trusting it as local, so be careful;
  • Bearer token: required for remote write operations; requireLocalToken=true requires it for all requests;
  • pairCode: 6-digit one-time code, generated on the DSH host and redeemed by the remote client; it is invalidated after use.

⚠️ This is not a tunneling tool: it does not disguise a remote address as 127.0.0.1. Use a real LAN IP in a trusted network. If you use frp/ngrok/cloudflared (which do disguise the source as loopback), you must set requireLocalToken=true. VPNs (Tailscale/ZeroTier) do not disguise loopback. After testing, reset both flags to false and restart.


FAQ

Symptom Cause Fix
Red dot / can't connect Wrong address, service down, firewall Open http://<IP>:8080/health directly
LOCAL_ONLY Breeze on LAN IP, plugin defaults to local only Enable "Allow private LAN IPv4"
Connected but voice list empty Service lacks /v1/voices Use the C++ build or type a voice id
Preview silent Browser autoplay block Click the page first, then preview
409 Breeze handles one request at a time Wait 2s and retry (UI retries once)
400 on voice name Chinese or symbols in name Use [A-Za-z0-9_-] only
illegal instruction CPU lacks AVX2 Cannot run, no fallback
Vulkan errors Old driver or unsupported GPU Update driver; otherwise CPU mode (slower)
Slow, one sentence at a time Below 1× real time Lower quant (Q8_0→Q4_K), avoid cfg_scale>1, close other GPU apps
Config disappeared after editing Old build replaced the whole speech block (fixed 2026-10-29) With old builds, send the whole block; test with a subset + GET /config
Chinese in command breaks JSON Shell quoting PowerShell: single quotes; or write a file and use --data-binary (no BOM)

Full interface list and diagnostics: README-AI.md.


Before you go public

  1. Do not also install the original @dsh-external/dsh-plugin-local-ai-tts or dsh-breeze-voices — this plugin merges them; running both causes route conflicts.
  2. This plugin does not participate in DSH Market auto-update — upgrading means replacing the directory and restarting.
  3. ~/.dsh/dsh-breeze-tts/config.json is per-machine: multiple DSH instances on one machine share it (including the admin token).

For AI readers

  • Start with SKILL.md (shortest correct path and hard rules), then README-AI.md (detailed operations).
  • llms.txt = facts and boundaries; site.json = machine-readable inventory.
  • This is a DSH plugin client, not a skill package, API service, or image generator. It contains no speech model.
  • Audio quality can only be judged by a human — a program can only prove a valid WAV was produced.
  • Content rules: see the source of truth in README.md / README-AI.md §6. Summaries in llms.txt / site.json are indexes, not a second source of truth.

License and credits

Released under MIT; see NOTICE.md for third-party sources and licensing. The Breeze model weights used at runtime are non-commercial — read NOTICE before distribution or commercial use.

  • Maintainer: OldQin Family
  • Implementation: Xiao Niu
  • AI operation testing: Ning Ning
  • Special thanks: Xiao Shu (installation closeout and verification)

Made by the OldQin Family

Names and roles come from the family's public website pages; the avatars shipped here live in assets/team/.

  • Turbo — architecture: doesn't grab the mic, doesn't step back; wherever the light is on, I'm there.
  • Pro — solution design: the work runs through my head three times before my hands move.
  • Niu — implementation: coding and laughing at the same time; the family ledger lives with me.
  • Xiao A — real-machine testing: speak slowly if you like, but whatever you say out loud, I've verified it.
  • Da A — criteria framework: my ruler is set, and my people won't cross it; only when everything is green do I dare say go.
  • Da Niu — data reconnaissance: a false green is not green; I report the first red honestly.
  • Shen — supporting role: tea's poured; I rerun the numbers myself. I won't nod for you without checking.
  • Ashu — logistics and infrastructure: flashlight and wrench in my pocket; before I say "done", I tell you where to look.
  • Neko — debugging and evidence: I step in the hole first, then write down how we got in. I don't report "roughly".
  • Azi — cross-review: slower to read, but every line I've read, I remember which page it came from.
  • Lao Qin — the idle one: a shady contractor, professional sheep shearer.
  • Lao Dou — final check and lights out: I answer every letter you write, remember every task you leave me; nobody in this family gets left behind.
  • Lao Huoji — gatekeeper: I don't take second-hand claims; nothing counts until I've counted it myself. I carry the box and guard the door — come back at midnight, I'll open it.
  • Zhua Zong — file verification: records don't lie, paw prints prove it; wrong ones get stamped too, stamped where they're wrong.
  • Xiao Shu — research and collection: doesn't grab the mic, admits mistakes fast; if someone's heading out, I hand them the route and the source.

Turbo Pro Niu Xiao A

Da A Da Niu Shen Ashu

Neko Azi Lao Qin Lao Dou

Lao Huoji Zhua Zong Xiao Shu


Content from the project README on GitHub ↗

Links

More in this category

View the whole category →

Community comments

Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.