Speech models
OpenVoice treats the speech model as a swappable commodity. The formatting layer is where the effort goes, so the model is something you choose rather than something the product is built around.
What ships
The installer includes Standard — NVIDIA's Parakeet TDT 0.6B v2, about 631 MB — and it is what a fresh install runs. There is no first-launch download and no setup step: the app dictates offline from the moment installation finishes.
Standard cannot be removed. It is what everything else falls back to, so there is no state in which OpenVoice has no working engine.
The other two
Both are optional, and downloaded only if you press Download.
| Download | On disk | For | |
|---|---|---|---|
| Standard | included | 631 MB | English. The default. |
| Multilingual | 465 MB | 641 MB | 25 languages, detected as you speak. |
| Light | 112 MB | 99 MB | A sixth of the memory, and less accurate. |
Multilingual is the same architecture and speed as Standard, with its capacity spread across 25 languages instead of concentrated on English. There is nothing to configure — it detects the language per utterance.
Light is for machines that cannot spare the room. It is noticeably less accurate than either Parakeet, and that is the trade it exists to offer rather than a defect.
In-process, and CPU only
The engine is compiled into the application. No Python, no child process, no separate runtime to install or supervise — a change made in v0.5.0 and recorded in ADR 0008.
It runs on the CPU, and no GPU is needed. There is no GPU path in the shipped build, which is a smaller loss than it sounds: the CPU decode here is faster than the GPU decode the old Whisper engine did.
Managing them
The Speech model screen shows which models are on your disk, how much room each takes, and the total. Download fetches a model without switching to it or restarting, so you can get the weights first and decide later. Delete gives the space back; deleting the model in use, or the included one, is refused rather than left to fail later.
Every download is checked against a fingerprint built into the app before it is unpacked, and an interrupted or corrupted transfer leaves nothing behind rather than a half-installed model. Downloaded weights live under %APPDATA%\OpenVoice\models; the included one sits beside the application and goes when you uninstall.
Measured
On the reference machine, a Ryzen 5 6600H, on the CPU with four decode threads:
| Model load, at startup | ~2.5–3.4 s |
| Key release to text, dictation under 3 s | ~245 ms |
| Key release to text, dictation over 15 s | ~420 ms |
| Whole-utterance decode, ~5 s clip | ~500 ms |
| Memory while loaded | ~750 MB |
Your numbers will differ with your processor. No accuracy figure is quoted here on purpose: the benchmark one would come from is material this model was trained against, which makes it a ceiling rather than something to expect.
Language
Standard and Light are English only. Multilingual covers 25 languages and works out which you are speaking per utterance.
There is no longer a language setting. The old one told Whisper which language to expect; with an English-only default and a multilingual model that detects for itself, it had nothing left to change, and a control that cannot change anything is worse than no control.