SoloSpark Voice Studio
connecting…
Text to speech

Type Mongolian, hear it in Namsrai's voice

0 / 2000
A/B compare
default
QA retakes

This session's takes

kept in memory until you close the tab (last 30)

    Nothing yet. Speak something and every take lands here.

    Training history

    How this voice was trained

    Every version is the same open model (OmniVoice, 0.6B) plus a small adapter trained on Namsrai's own speech: his published YouTube videos plus clean-mic recordings, about 0.4 h from 1.0 on, nearly doubled in 2.0 by a studio read-through (0.76 h). Each version was scored on the same 20 unseen sentences.

    clean-mic recordings his YouTube videos live streams

    Sounds like him higher is better

    ★stableexperimental side branch (off the main line)
    Speaker similarity (0 to 1): how close each version sounds to Namsrai's real recordings, measured by a speaker-recognition model. The dashed line is his own clean-mic clips.

    Clear to understand lower is better

    ★stableexperimental side branch (off the main line)
    Whisper CER: the share of letters a speech recogniser gets wrong when it listens to the voice and writes down what it heard. Lower means clearer speech.

    SoloSpark Voice · runs offline on this PC · voices are non-commercial (CC-BY-NC weights)