ζ—₯本θͺž TTS β€” piper-plus vs Style-BERT-VITS2 build 37

sets Pitch, Voice size and Breathiness reduction on both panels (chosen by ear for tsukuyomi-chan)
sentences with no Japanese at all; piper-plus reads them in its English mode (with a Japanese accent), Style-BERT-VITS2 with the browser's voice

piper-plus tsukuyomi-chan Β· via yomiage Β· 65 MB download, kept on the device

How fast the voice talks (1.0Γ— = normal, 1.5Γ— = 50% faster). Changes tempo only, not pitch.
How high or low the voice is, in semitones (12 = one octave). Applied after the voice is generated; speed and voice size stay the same. Negative = lower.
Apparent size of the speaker's throat and mouth, independent of pitch. Negative = bigger, deeper, more adult/male. Positive = smaller, younger, cuter.
Softens the hissy, airy high frequencies (above ~3 kHz). Higher = less breath and hiss, but consonants like s/sh get duller. Try 0.3–0.6.
Random variation in the voice's tone and pitch contour. Low = flatter and steadier, a bit robotic. High = livelier but can sound wobbly or unstable. piper-plus docs: "audio variation control"; they recommend 0.5 for Tsukuyomi (their CLI default is 0.4).
Random variation in timing: how long each sound and pause lasts. Low = metronome-steady. High = more human-sounding but uneven. piper-plus docs: "phoneme duration variation"; their CLI default is 0.5.

      


      
      

        
Style-BERT-VITS2 JP-Extra + DeBERTa Β· ~510 MB

        
        

          

            How fast the voice talks (1.0Γ— = normal, 1.5Γ— = 50% faster). Changes tempo only, not pitch.

          

            How high or low the voice is, in semitones (12 = one octave). Applied after the voice is generated; speed and voice size stay the same. Negative = lower.

          

            Apparent size of the speaker's throat and mouth, independent of pitch. Negative = bigger, deeper, more adult/male. Positive = smaller, younger, cuter.

          

            Softens the hissy, airy high frequencies (above ~3 kHz). Higher = less breath and hiss, but consonants like s/sh get duller. Try 0.3–0.6.

          

            Random variation in the voice's tone and pitch (the model's noise scale). Low = steady and flat. High = more expressive but less stable. Default 0.6.

          

            Random variation in how long each sound lasts (noise W, used by the stochastic duration predictor). Low = steady timing, high = more varied. Default 0.8.

          

            Mix between two timing models: 0 = always the deterministic one (even, predictable rhythm), 1 = always the stochastic one (varied, more natural-sounding rhythm). Default 0.2.

          
        

        

          
          
          
        

        

      

    


    
    

      
Reading view Sudachi Β· furigana Β· bunsetsu Β· ~125 MB, cached in IndexedDB

      

        
        Shows the text above with furigana on kanji and a space between bunsetsu. Hover or click a bunsetsu for its dictionary form.
        
      

      

        
      

      

        
      

      

        
      

      

        nounpronounverb
        い-adjectiveγͺ-adjectiveadverb
        pre-noun adjectivalconjunctioninterjection