Choose a TTS engine
Narrow VoiceStudio's TTS engines by cloning, language, hardware, and model terms.
VoiceStudio v0.5.0 lists 16 TTS engines. Choose by the job and the machine that will run it, then compare voices on the same script.
Source for this guide
The capabilities below come from the immutable v0.5.0 engine matrix. Language coverage does not guarantee equal quality for every language or speaker.
Does the job need voice cloning?
Start here for dubbing or a batch job that must preserve a speaker. The v0.5.0 matrix marks KittenTTS, Sherpa-ONNX, and Supertonic 3 as not supporting cloning. VoiceStudio rejects those choices for dubbing or a pinned batch voice instead of silently changing engines.
KittenTTS remains a direct CPU choice when an English preset voice is enough. For speaker preservation, keep only rows with cloning marked in the matrix.
Match the engine to the hardware
On Apple Silicon, MLX-Audio is the native adapter. Its cloning and instruction capabilities vary with the selected MLX model. The default VoiceStudio engine, CosyVoice 3, VoxCPM2, and the OmniVoice subprocess path also list MPS support.
CPU-only machines have several different options:
- KittenTTS handles English without cloning.
- Supertonic 3 lists 31 languages without cloning.
- OmniVoice GGUF lists 600-plus languages, cloning, and instruction control.
- PocketTTS lists six languages and cloning. Its model access is gated, so review its terms before installation.
CUDA supports the broadest set of rows in the matrix. Model size still matters. MOSS-TTS v1.5 and dots.tts are documented as heavyweight optional engines, so check their memory requirements before installing them.
Check language and instruction control
The default VoiceStudio engine has the widest listed language coverage and supports cloning plus instruction control. CosyVoice 3 and VoxCPM2 also list both capabilities with smaller language sets. GPT-SoVITS lists five languages and cloning without instruction control.
Test the exact language, speaker, and script you plan to use. A language count is coverage information, not a quality score.
Shortlist by task
| Need | First rows to inspect | Check before installing |
|---|---|---|
| Multilingual cloning and instruction control | VoiceStudio, CosyVoice 3, VoxCPM2, OmniVoice GGUF | Exact language, hardware path, model terms |
| English CPU output without cloning | KittenTTS | A preset voice is sufficient |
| CPU cloning in a six-language set | PocketTTS | Language, gated access, license conditions |
| Native Apple Silicon model adapters | MLX-Audio | Capabilities and license of the chosen model |
| Dubbing with a pinned speaker | Any row marked for cloning | Exclude KittenTTS, Sherpa-ONNX, and Supertonic 3 |
| Crash-isolated unattended synthesis | OmniVoice subprocess | Separate process and memory tradeoffs |
This table narrows the catalog. It is not a benchmark or a universal ranking.
Read model terms before installation
The app's own AGPL or paid commercial licence does not replace upstream model rights. Review the exact weights, selected voice, tokenizer and other required components. Commercial model use, commercial output use and redistribution are separate scopes; a permissive inference SDK label does not settle any of them.
The Library joins the existing engine catalogue to the app's versioned model-licensing export. Each model page shows declared terms, original evidence, configured and evidence revisions, dates, unresolved checks and the registry fingerprint. The JSON export preserves the complete source inventory and unresolved dynamic asset families.
This is a disclosure-only foundation. No complete real model workflow is verified ready in the current snapshot, and this website does not claim every released installer or SDK download already enforces review. The licensing export's app source pin is separate from the older capability snapshot and may be newer than your installed release. Review the matching app version before relying on it.
“Not verified” is neither permission nor a prohibition. A known restriction is shown separately. A local acknowledgement cannot replace an upstream agreement, provider gated-access approval, a separate commercial grant or voice consent. Full original upstream terms govern; summaries do not rewrite them.
Read the app and model licensing guide, shortlist only the engine you plan to test, and compare it with the default using the same inputs.