Ming
inclusionAI's Ming-3.0-tts run locally in Voxta. Cloned, trained and described voices, with emotions and inline non-verbal cues. CUDA only.
Ming runs inclusionAI's Ming-3.0-tts (0.5B) locally. It is light, using about 3 GB of video memory. It picks a voice three ways: cloned from a short sample, one of the voices it was trained on, or described in plain words.
Experimental and admin-only. Runs on a CUDA GPU only, and is barely faster than real time.
Setup
Add the service
Add Modules → Services → Ming → Install, then follow the install steps. The model downloads in the Models step. The install brings in the Python runtime and, on Windows, FFmpeg (watch the Terminal for progress).
Pick a voice
In Choose a voice (on the character's Voice tab), pick a sample from the voice library, the character's own voice sample, or one of the voices listed as Trained. Default Female / Male Voice in the module settings cover characters without one.
Three kinds of voice
| Kind | How | Emotions |
|---|---|---|
| Cloned | A voice sample, with its transcript. Keep it short: samples longer than Max Voice Sample Length (default 6 s) are refused, since a long one produces noise. | None. A clone ignores direction, so affect clips are not used. |
| Trained | A voice Ming was trained on, picked in Trained Voice. | The ones it was trained with, shown in the picker. Most have happy, sad, angry, fearful, disgusted and surprised. The English ones have none. |
| Described | Voice Description in plain words, e.g. "A young woman with a warm, slightly husky voice, speaking calmly." Takes precedence over the sample and the trained voice. | Any word in Described Voice Affects (happy, sad, angry, whispering, playful, sarcastic...), added to the description. |
A described voice is drawn anew for each line, so keep its Seed fixed to keep it sounding like the same person.
Languages: English, Chinese, Japanese, Korean, German, French and Spanish. Auto uses the chat's language when Ming speaks it.
Inline cues
The Context tells the AI it may write <laugh>, <sigh>, <cough>, <breath>, <surprise> and <uhm>. Speech Replacements turn *laughs*, *gasps* and the like into them, and drop any other tag.
Settings
| Setting | Notes |
|---|---|
| Guidance Scale | 1–4, default 2. Higher follows the text and voice more strictly. |
| Sigma, Temperature (advanced) | The sampler's noise and randomness. Defaults 0.25 and 0. |
| Thinking Speech (advanced) | Short sounds played while the reply is being generated. |