Engine
Studio: the two voices of Matcha-TTS-PL-v2, each with its large vocoder.
chunk (frames)
ODE context
vocoder context
first chunk
Defaults (43 / 16 / 12) = the sweet spot of a 210-setting sweep.
Voice
training dataset voices and custom mix
These are the audiobook readers the model was trained on. Their voices are theirs: see the licence note below before using a recognisable reader. Moving a slider makes a custom mix.
advanced: durations, pitch range, temperature, tempo, English words, ODE steps, backend
temperature
tempo (length scale)
pronunciation dictionary
English words (meeting, feedback, YouTube…) are detected and pronounced in English automatically (switch above); mark any other word as English with braces:
{Hugging Face}. Rules here: one per line, word -> how to say it in Polish spelling, whole words, case-insensitive, applied first. Saved in this browser.loading…
About the voices. Matcha-TTS-PL v2 has two synthetic voices (male and female), one model each with its own vocoder: designed from text descriptions with VoxCPM2 (OpenBMB, Apache-2.0), not recordings or clones of real people. Each model was distilled from about 11 h of its voice reading conversational Polish (what a robot says to a person: reactions, questions, short and long turns, plus assistant lines and YouTube transcripts), English words included, and speaks with stochastic durations by default: a flow-matching duration predictor draws the rhythm anew every time (switch it off under advanced). Switching voices downloads that voice's model (about 140 MB). Weights: CC BY-SA 4.0; every text source is credited in the model's ATTRIBUTION.md. Training data: polish-synthetic-speech-voxcpm2-v2. Playground code MIT; espeak-ng GPL-3.0.
Fine-tuned from the pretraining checkpoint Matcha-TTS-PL-Pretrain (audiobook readers; not intended for production).
History (newest first). Each clip keeps its settings; send it back to the effects panel with “→ effects”.
Effects
Effects apply to the clip playing in History (the newest one until you play another). Its ⬇ wav saves what you hear.