Kyutai's Pocket TTS clones a voice from 5 seconds of audio, on CPU, under MIT. Benchmarked against Kokoro, Supertonic, and Inflect-Nano for Eng. TTS

Kyutai dropped Pocket TTS a bit ago and I've been sitting on it for a benchmark. Finally ran it head to head against the three CPU TTS models that have been getting attention (Kokoro 82M, Supertonic 3, Inflect-Nano-v1). 180 timed runs, 36 audio samples, objective MOS scores via UTMOS.

Short version: Pocket TTS is the slowest of the six configs I tested, and it's still the most interesting model in the field. Here's why.

It's a ~100M param streaming language model that generates audio tokens over Kyutai's Mimi neural codec, then decodes to 24kHz. So instead of the usual acoustic-model-plus-vocoder setup, it's more like an autoregressive LLM but for audio. Token by token.

- Latency is dead flat across text lengths. Its RTF is 0.69 to 0.76 whether you feed it 12 chars or 1712 chars. No fixed overhead to amortize. Compare with Kokoro PyTorch which climbs from 0.49 on tiny text to 0.83 on long text.

This is the headline feature. Hand it a 5-second reference clip of any voice and it speaks in that voice. Accent, timbre, pacing, even the mic character of the reference. No fine-tuning. No GPU. MIT license.

None of the other CPU-friendly models can do this at all. Kokoro and Inflect-Nano ship fixed voice sets, Supertonic same. If you want a user-supplied voice on a CPU box, Pocket TTS is currently in a category of one.

I ran the benchmark with Pocket TTS pinned to a preset voice (alba) for a fair speed/quality comparison. The cloning capability isn't in the numbers below because you can't benchmark it against models that don't have it.

Config Mean RTF UTMOS MOS Params License Supertonic 3 (2-step) 0.121 1.53 ~99M OpenRAIL-M Inflect-Nano-v1 0.145 3.48* 4.6M Apache 2.0 Supertonic 3 (5-step) 0.240 4.32 ~99M OpenRAIL-M Kokoro 82M (ONNX) 0.641 4.44 82M Apache 2.0 Kokoro 82M (PyTorch) 0.665 4.46 82M Apache 2.0 Pocket TTS 0.714 4.10 ~100M MIT Hardware: Intel Xeon 8272CL, 4 cores, 16GB RAM, no GPU. UTMOS is utmos22_strong, an objective MOS predictor, so it's not just my ears this time.

The Inflect-Nano asterisk: UTMOS gave it 3.48 but to the ear it's buzzy and robotic. Known UTMOS failure mode where it over-rates small HiFi-GAN vocoders for being clean rather than natural. Also it has a hard ~15 second output cap I discovered mid-benchmark, so its RTF on long inputs is inflated.

- Tiny footprint for short utterances → Inflect-Nano-v1, if you can live with the buzz and the 15s cap

Pocket TTS install is genuinely painless. pip install pocket-tts, no CUDA build, no HuggingFace-repo-plus-sys.path wiring. Downloads weights on first load. The least fussy of the six.

The MIT license is a big deal. Kokoro is Apache 2.0 (also great). Supertonic is OpenRAIL-M with commercial restrictions. Pocket TTS being MIT means you can do essentially whatever with it commercially.

Repo with raw CSV (180 rows), all 36 WAV samples, and the benchmark script is in comments below 👇

If anyone here has run Pocket TTS voice cloning with a real reference clip, would love to hear how it holds up on different voice types (accented English, non-English, singing, etc). That's the next thing I want to test but I need a clean dataset.