Choosing a model

Start with loudr-1 for reference quality. Try loudr-1-turbo when you want audio sooner. Compare them on your own text and voice before deciding: they produce different readings, even with the same seed.

Switch by changing the name. Keep the rest of your code and your voice profile:

import loudkit as lk

engine = lk.load("loudreader/loudr-1")
# Or: engine = lk.load("loudreader/loudr-1-turbo")
voice = engine.voice("joe")
engine.synthesize("Hello from loudkit.", voice, seed=7).save("hello.wav")

Where they run

Both models are supported by the 0.1.1 code:

Your environment	Runtime	loudr-1	loudr-1-turbo
Python on CPU or GPU	PyTorch	yes	yes
Python without PyTorch	ONNX Runtime	yes	yes
Python on a Mac	CoreML	yes	yes
Swift on macOS or iOS	native generator + CoreML renderer	yes	yes
Go, Rust, JavaScript / TypeScript	ONNX Runtime	yes	yes

Use a release containing the graphs for your runtime. See troubleshooting if a download or model load fails.

What changes

Turbo generates pairs of speech tokens and uses one renderer estimation step. It does less model work, but the speedup depends on the device and passage. Benchmarks distinguishes measured results from expectations.

Listen for naturalness, pauses and pronunciation on the material you plan to read. A faster result is useful only if you like listening to it.

What stays the same

The API, available voices, text preparation, streaming and voice-profile format are shared. Clone a voice once and reuse its file with either model. Loading an existing profile does not load the enrollment models.

A seed makes a reading repeatable within the same model and execution setup; it does not make the two models produce identical audio.

Next

Getting started · Clone a voice · loudr-1 model card · Turbo model card