OmniVoice
Text to speech, voice design from a description, and voice cloning from a few seconds of audio.
- Powers
- Voice Clone, Text to Speech
- Modes
- Text to speech · Voice clone
- GPU class
- 24 GB to 48 GB Ada
- In a fresh install
- Off; one switch in Settings → Models
01
What it is
OmniVoice is an open-weight speech model that does three things: reads text in a natural voice, designs a voice from a plain description ("a warm, low male voice, slightly hoarse"), and clones a voice from a short sample. Whisper transcribes the sample so the clone keeps the speaker's rhythm as well as the timbre.
It powers Text to Speech and Voice Clone. A sample can be recorded in the app (up to 15 seconds) or uploaded as WAV, MP3 or M4A; the first 20 seconds are used. Long texts are read in chunks of about a hundred words and joined.
02
Tools it powers
- Voice CloneA short sample of a voice, then any line in that voice.
- Text to SpeechNatural speech from text, with voice design: describe the voice you want.
03
Modes
- Text to speechText in, speech out; an optional voice description designs the voice.
- Voice cloneA sample plus the text to say in that voice.
04
Settings exposed
- SpeedPlayback pace, default 1.0.
- Voice designA description of the voice (text to speech).
- Reference textWhat the sample says, if Whisper's transcript needs correcting (voice clone).
- Steps and guidance4 to 64 steps, default 32; guidance scale default 2.0.
- Words per chunkHow long texts are split before they are joined; default 100.
- Clean-upDenoise the sample and post-process the output, both on by default; t-shift, position, class and layer-penalty controls for the curious.
- SeedRandom, or fixed.
- Custom LoRAsNot applicable.