Kokoro-82M became one of the most-used open text-to-speech models because it’s tiny, fast and sounds good. Its limitation has always been that you get the voices it ships with. Inno Clone Tuner, from developer remsky (who also maintains the popular Kokoro-FastAPI server), changes that. Its downloads on Hugging Face rose 38% in a single day in our tracker, to about 153,000.
What it does
You give it a reference clip: at least 3 seconds, and it only reads the first 30. It returns a standard Kokoro voice pack, the same [510, 1, 256] tensor the built-in voices use. No fine-tuning happens, and Kokoro itself isn’t modified. The new pack just drops in.
The adapter is small: 9.65M parameters, about 24MB at fp16. Enrollment takes about 0.05 seconds per second of reference audio on a CPU, or 0.1–0.3 seconds total on a GPU. It installs with pip install inno-kokoro and ships in Kokoro-FastAPI v0.9.0+ behind a single environment flag.
How it works
There are three parts:
- A distilled WeSpeaker ResNet34 speaker encoder (6.6M parameters) that listens to the reference and produces a speaker embedding. In a Sept 1 update, remsky retrained the distillation on real archival radio recordings to close a quality gap on older, noisier audio.
- A style head that maps the embedding into Kokoro’s 256-dimension timbre space. It also shifts the result along a learned direction that tracks spectral tilt, meaning how bright or dark the reference sounds.
- A prosody blend. Rather than inventing a new speaking style, it measures the reference’s average pitch, pitch range and syllable rate, then builds a weighted mix of Kokoro’s existing voices that best matches those numbers.
That third part explains both the tool’s strength and its ceiling. Rhythm and intonation always come from voices Kokoro already speaks well, so the output rarely goes wrong. But the output never really picks up a speaker’s own delivery either.
The honest benchmark
The README includes an identity benchmark on LibriSpeech test-clean (1,127 utterances, 39 held-out speakers). It’s more candid than most model cards. On a scale where 0 sounds like a stranger and 1 sounds like a second recording of the same person:
- F5-TTS v1: 0.94 identity, 3.86 quality (UTMOS), slowest at 0.48 real-time factor
- StyleTTS2 zero-shot: 0.46 identity, 4.40 quality
- Inno v0.2: 0.32 identity, 4.45 quality, 0.07 real-time factor
- OpenVoice v2: 0.23 identity, 3.80 quality
- Kokoro’s nearest stock voice: 0.15
In plain terms, Inno gets you about a third of the way to someone’s voice. It has the highest quality score in the table, including above the real recordings (4.10), because it won’t copy room noise and microphone artifacts. remsky states the trade openly: “some texture/identity is traded to maintain audio quality and stability.”
Who it’s for
If you need a convincing replica of a specific person, this isn’t the tool, and F5-TTS is far ahead. If you need a family of distinct, clean, fast voices for a game, installation, prototype or narration project, Inno is well suited. Point it at a few permissioned recordings and you get voices that are in the right neighborhood and sound polished at Kokoro’s speed. It’s English-only for now.
That shallow clone also lowers the risk of impersonation, though it doesn’t remove it. The README’s closing line applies regardless: only clone voices you have permission to clone. Licensing is Apache-2.0, except the built-in speaker encoder, which is CC BY-SA 3.0. Check that before shipping it in a product.