OmniVoice
A diffusion language model that speaks over 600 languages and designs voices from a description.
k2-fsa/OmniVoice on Hugging Face“Speaking is three to four times faster than typing for most people. With local dictation, you hold a key, say what you mean, and a small model on your own machine turns it into text. Nothing leaves your computer, there is no subscription, and it keeps working on a plane. The cleanup step removes filler words, fixes punctuation, and formats lists, so what lands in your editor reads like you typed it carefully.”
Memory
The most it held at once, from loading to the last test. Next to every other text to speech model here.
Speed
How long it takes to make speech, from the text going in to the last sample coming out.
Clarity
Twenty sentences full of numbers, times, addresses and codes. Whisper large-v3 turbo listened to each, and both sides were normalized, so “5” and “five” count as the same.
- 01
The quick brown fox jumps over the lazy dog.
“The quick brown fox jumps over the lazy dog.”· heard word for word
0:03 - 02
Please call me back at 415-555-0132 before five o'clock.
“Please call me back at 415-555-0132 before 5 o'clock.”· heard word for word
0:05 - 03
Send the invoice to billing@localflow.dev by Friday.
“Send the invoice to billing at localflow.dev by Friday.”· 1 word off
0:03 - 04
Our office is at 1600 Amphitheatre Parkway, Mountain View, CA 94043.
“Our office is at 1600 Amphitheatre Parkway, Mountain View, CA 9443,”· 1 word off
0:06 - 05
The meeting moved from 3:30 PM to 4:15 PM on March 3rd.
“The meeting moved from 3:30 p.m. to 4:15 p.m. on March 3rd.”· heard word for word
0:05 - 06
It costs $42.99, or about 39 euros with tax.
“It costs 42.99 or about 39 euros with tax.”· 1 word off
0:04
How it ran
Ran with audio.cpp on Vulkan, q8_0 weights, voice designed as “female, young adult, moderate pitch” on an AMD Radeon RX 9070 XT (16 GB, Vulkan through RADV (Mesa 26.2.2)).