Generating a voice
outpaces speaking.
Ocelot is a small model distilled from the speech synthesis model Qwen3-TTS.
Ocelot is a small text-to-speech model that turns text into spoken audio. Press the button: the upper tape advances at the actual generation time, and the lower tape advances at the speed of the audio.
—x faster
Small enough
to run anywhere.
Ocelot is made of two small networks: one that turns text into a blueprint of the voice (a mel spectrogram), and one that turns the blueprint into a sound wave. As you scroll, the view zooms out and the difference in size from other models comes into view.
* Speed on smartphones is still being measured. provisional
Listen
to 20 languages.
We had Ocelot read a short, a medium and a long text with the same content in every language. Press a waveform to play it. You can choose from three voices (more voices are on the way). During playback, the movement of the voice's pitch is drawn over the waveform as a line.
Press a waveform and the text being read appears here.
* Everything here is audio actually generated on a CPU when the figures on this page were measured, exactly as it came (unedited). The content of the texts is fictional.
Have it read
your own text.
Type a text, choose the language and the voice, and press the button. The audio is generated on the spot, on this server's CPU. Up to 200 characters.
* The language is not detected automatically: choose the language the text is written in. Numbers, abbreviations and foreign proper names may be misread.
Long texts,
at the same speed.
For all 180 clips: the length of the audio (horizontal) and the time it took to generate (vertical). Every point is below the line marking one tenth of the speaking time.
Changing the number of CPU threads
Does a machine
understand it, too?
This is the character error rate: we turned the generated audio back into text with speech recognition (Whisper large-v3-turbo) and compared it with the original text. The lower the number, the more clearly the text was read.
From text
to voice.
The four stages, shown with one actual sentence. As you scroll, the shape changes in order from the left.
Text
Phonetic symbols
Voice blueprint
Sound wave
- Conversion to phonetic symbols uses dedicated dictionaries for Japanese, English and Croatian, and espeak-ng for the other 17 languages.
- The language is passed to the model as a "speaker ID." A single model reads all 20 languages.
- Long texts are split into sentences, generated one sentence at a time, and joined with 0.25-second pauses in between.