01 / Speech synthesis

RESONANCE

A scratch-trained speech model you can listen to.

Six saved phrases · zero inference cost per visit

Speech desk / six saved generations

Give the voice a line.

Type a recorded line or choose one below. Hear its output from the trained model, generated locally and saved for this exhibit.

Your line

Select or type one of the six saved lines. Your text stays in this browser.

Output waveform0:00

Loading the recorded phrase collection…

Generated from acoustic checkpoint step 17,500 and vocoder checkpoint step 20,000.

Reading the signal

Text

A character sequence enters the acoustic model. The six phrases here were generated from the saved checkpoint.

Sound

The model predicts 80-band mel features; a vocoder turns them into the waveform you hear.

Plot

The training figure tracks mel error over optimizer updates. Listening reveals timing and pronunciation details that one loss value cannot describe.

RESONANCE acoustic training loss, saved waveforms, and run measurements — This saved figure tracks acoustic loss over 10,000 optimizer updates and compares sample waveforms. The listening desk above uses a later selected checkpoint.

RESONANCE maps text to a mel spectrogram, then turns that spectrogram into audio. Its acoustic and waveform models were trained from random initialization on LJ Speech.

The listening desk contains six saved generations. The waveform shows each clip's amplitude over time; the training figure above shows the acoustic model's measured learning curve.

See the local observatory

A dark speech model observatory showing spectra, learning curves, and signal measurements