loudr-1-turbo

The faster of the two loudkit models. Same 28 voices, same ten languages, same API.

loudr-1-turbo is loudr-1 with two changes to how the audio is produced, and no change to how it is used:

• the token generator writes two speech tokens per forward instead of one, so a second of audio costs half as many;
• the renderer turns those tokens into audio in one pass instead of several.

Everything else is loudr-1: the same voices, the same text handling, the same sampling, the same seeds.

Which one to pick

loudr-1 is the default. Pick it when you want the reference quality.

loudr-1-turbo is faster on the same hardware, at a small cost in naturalness that is easiest to hear on long, quiet or heavily punctuated passages. Pick it for interactive reading, where latency is what the listener notices. Listen to both on your own text before choosing, and measure both on your own hardware; the benchmark page has the commands.

Start in Python

pip install "loudkit[torch,audio,hub]"

import loudkit as lk

engine = lk.load("loudreader/loudr-1-turbo")
voice = engine.voice("joe")

engine.synthesize("Hello from loudkit.", voice, seed=7).save("hello.wav")

That is the whole difference from loudr-1: one string. The same from a shell:

loudkit speak --checkpoint loudreader/loudr-1-turbo --voice joe \
  "Hello from loudkit." -o hello.wav

To enroll a voice that you own or have permission to use:

mine = lk.enroll("my-recording.wav", "loudreader/loudr-1-turbo", name="my-voice")
mine.save("voices/my-voice.safetensors")

A voice profile is interchangeable between the two models: a profile enrolled against loudr-1 loads against loudr-1-turbo and the other way round.

Backends

Version 0.1.1 supports both models in Python, Swift, Go, Rust and TypeScript. Python offers PyTorch, ONNX Runtime and CoreML; Swift uses its native token generator with CoreML rendering, and Go, Rust and TypeScript use ONNX Runtime. Changing the model name keeps the same synthesis API and voice profile.

Download

loudkit download loudreader/loudr-1-turbo --for onnx

Choose --for torch or --for coreml for another backend. Add --with-cloning to prepare enrollment as well. Loading an existing voice profile does not require enrollment assets. Prepare the chosen set once to use it offline.

What ships

The bundle uses the same canonical enrollment weights and graphs as loudr-1. Clone once, then use the unchanged profile with either model. Exact file sizes and checksums are recorded in the bundle's release.json and SHA256SUMS.

Listen

These samples use this model, the named shipped voice and seed 7.

Voices and consent

The 28 profiles are the ones loudr-1 ships, unchanged: ten for English and two each for Spanish, French, German, Italian, Polish, Portuguese, Dutch, Swedish and Danish, built from recordings donated for speech technology or from CC0 and CC-BY speech corpora. No scraped celebrity voices ship with the model. The full roster records the source, licence and consent basis for every profile.

Model lineage

loudr-1-turbo is derived from loudr-1, which is derived from Chatterbox, released by Resemble AI under the MIT licence. The token generator is a student of loudr-1's, trained to emit two tokens per forward; the renderer is loudr-1's, distilled to reach the same audio in one pass. The licence chain is loudr-1's, unchanged, and is recorded in NOTICE.

How the two-token generator works is written up in two-token decode.

Reproducibility

For a fixed build, device and backend, the same text, voice and seed produce the same waveform. loudr-1 and loudr-1-turbo are different models: the same text, voice and seed give different audio on each, and a saved WAV records which one spoke. The contract is in the identity contract.

Before you ship

• Long passages are rendered in windows of about ten seconds. Sentence joins can occasionally be audible, and turbo joins a little more so than loudr-1.
• Difficult punctuation, numbers and abbreviations can change pronunciation or prosody.
• Voice cloning requires consent. A recording being public does not grant permission to clone the speaker.
• Saved WAVs and server responses carry an unsigned, machine-readable note by default, recording the model, voice, seed, backend and a checksum of the audio. It is loudkit's own and is not C2PA Content Credentials.

Read Responsible use before exposing enrollment to other people.

Intended use

loudr-1-turbo is intended for local narration, accessibility, localisation, games, prototyping and speech research. It is not a voice-authentication system and must not be used for deceptive impersonation.

Training data

The original Chatterbox training data is controlled by Resemble AI and is not documented by this project. The shipped voice profiles use recordings made or released for speech-technology use; their sources and licences are listed in the public roster.

Licence

Apache-2.0. Upstream attributions and component licences are listed in NOTICE.

Download sizes for 0.1.1

Approximate decimal sizes for the release files; backend weights are included. Cloning adds enrollment assets only when requested. Both models ship separate synthesis and enrollment checkpoints.