dsh-dania-voice

dsh-dania-voice

English | 简体中文

Local GPT-SoVITS text-to-speech for DeepSeek Harness. It reads assistant replies aloud in a fine-tuned voice, controlled by a small draggable mascot hanging at the top of the interface.

Everything runs on your machine — no audio, no text leaves it.

How it works

DSH session events (assistant/message + turn/end)
      │  take the reply text: inline code → speakable English,
      │  code blocks → one short cue, tables and links stripped
      ▼
Host: clean + merge into ≤90-char segments
      │  POST http://127.0.0.1:9880/tts
      ▼
GPT-SoVITS api_v2.py (local, GPU)
      │  wav → cache dir → queue
      ▼
Browser: poll /dania-voice/next → Web Audio playback

Playback happens in the browser on purpose: the Host process has no audio device, and the voice should come out of the machine you are actually sitting at.

When cleaning, inline code is not dropped — ops_bash is spoken as ops bash, and long all-caps words are lowercased so the model reads them as words instead of spelling out E X E C U T O R. Code blocks become a single spoken cue ("there's a code block here") instead of being read line by line.

Requirements

Item	Note
DSH	>= 0.1.7-rc.1
Python	3.10 – 3.12 (the installer looks for it; it does not install Python)
GPU	NVIDIA with 8 GB+ VRAM recommended. CPU works but is several times slower.
Disk	~5 GB (PyTorch + base models + voice model)

Install

dsh plugin --profile web add dsh-dania-voice

Then open the mascot panel and click One-click setup. The installer runs as a separate process and reports progress into the same panel:

1. creates a Python venv under the install directory
2. installs PyTorch (CUDA build when nvidia-smi is present)
3. installs the inference dependencies
4. downloads GPT-SoVITS source (~50 MB)
5. downloads the pretrained base models (~1.6 GB)
6. downloads the voice model and its reference audio (~320 MB)

It is idempotent — if it fails halfway, click again and completed steps are skipped. That matters because step 5 is a 1.6 GB download.

The default install directory is $DSH_HOME/dania-voice. Override it with installDir if you want it on another drive.

Usage

A mascot hangs at the top centre of the window:

• Click — toggle reading on/off
• Drag — move it anywhere; the position is remembered
• Right-click / Shift+click — open the settings panel
• While reading, the mascot gently bobs

Panel controls: volume, speed, temperature (lower = steadier, less breathiness), inter-sentence pause, plus repeat last reply and say a line for testing.

Changing speed/temperature applies to the next synthesis; audio is cached by text + parameters, so repeating with the same settings is instant.

Configuration

Written to $DSH_HOME/dania-voice/config.json by the panel. The config: block in your profile patch works too and is overridden by the panel.

Key	Default	Meaning
enabled	true	Master switch
volume / speed / temperature	1 / 1 / 0.7	Playback and sampling
speakUser	false	Also read your own messages
installDir	$DSH_HOME/dania-voice	Where the engine lives
pythonPath / root / cacheDir	derived	Only needed for a hand-made install
baseUrl	http://127.0.0.1:9880	GPT-SoVITS endpoint
autostart	true	Start the service when needed

Known limitations

Short English tokens sound mushy. A full English sentence comes out fine — transcribing a synthesised sentence back gives an almost exact match. But short identifiers or ones with digits (ops bash, cut2, sha1) come out close to Chinese syllables, as if the voice were mumbling. That is a mismatch between the voice model and English phonemes, not something the plugin can fix. Inline code is still turned into speakable English; longer identifiers fare better than short ones.

A segment is capped at 90 characters. The server's generation loop is hard-coded to 1500 semantic tokens (~30 s); past that it stops silently and truncates. With 147 characters in one unsegmented request, a whole sentence went missing in testing. 90 leaves about 30% headroom, at the cost of more pauses between segments.

URLs and tables are not spoken. A URL reads as a string of letters nobody can follow, and a table reads as a list of numbers; both are dropped. The dropped spot is closed with a full stop so the neighbouring sentences are not glued together.

Troubleshooting

• Mascot shows a red border — the plugin's routes are not answering; check that the plugin is enabled.
• "Not installed yet" — click one-click setup, or check $DSH_HOME/dania-voice/install.log.
• No sound but the panel says ready — the browser blocked autoplay; click the mascot once.
• Slow first line — the model takes 20–40 s to load into VRAM; later lines take 1–3 s.

Credits

• GPT-SoVITS by RVC-Boss — MIT.
• Dania voice model by villia on ModelScope. The installer downloads it from the original repository; it is not redistributed here.
• Mascot artwork © 夜游人 (this project's author). Background art uses official promotional material under the publisher's third-party usage policy.

This is a fan project. Wuthering Waves and its characters belong to Kuro Games.