nano-edge-tts

A 1.75M-parameter decoder for Kokoro-82M that runs no neural network at audio rate.

Kokoro's text side (pronunciation, timing, pitch and all voices) is unchanged. Its 53.3M-parameter waveform decoder is replaced by two small networks that run 80 and 160 times a second and set per-frequency gains on a pulse train and on white noise. Two inverse STFTs turn the result into audio.

Paper: No Network at Audio Rate: A 1.75M-Parameter Harmonic-plus-Noise Decoder for Kokoro-82M (method, results, listening tests and negative results).

	Kokoro-82M	nano-edge-tts
decoder parameters	53.3M	1.75M
whole model parameters	81.8M	30.2M
decoder weights on disk	213 MB	7.0 MB
decoder speed, one CPU thread (real-time factor)¹	0.1378	0.0042 (33× faster)
phonemes to audio, one CPU thread (real-time factor)²	0.1411	0.0127 (11× faster)
peak memory, whole pipeline²	1.4 to 1.7 GB	0.75 GB
Whisper small.en word errors, 240 sentences	235 (4.63%)	238 (4.69%)
UTMOS, 28 validation clips	4.383	4.327

¹ 240 validation sentences, PyTorch, Apple M4. ² Six validation sentences, 55 s of audio, same machine. Raw numbers are in results/.

How it sounds

The same sentence through Kokoro's decoder and through this one:

voice	Kokoro-82M	nano-edge-tts
af_heart
af_bella
am_michael
am_fenrir
bf_emma
bm_george

In blind A/B tests by one listener (the author), an earlier checkpoint was heard as the same as Kokoro on 9 of 12 pairs. The released model is close to Kokoro on most sentences and not identical: on some, voiced sounds between 1 and 4 kHz are a little too pure, heard as a faint machine-like tone. The paper describes the defect and two attempts to train it away that did not work.

Quick start

pip install kokoro==0.9.4 safetensors soundfile
hf download drmhse/nano-edge-tts --local-dir nano-edge-tts --exclude "samples/*"
cd nano-edge-tts

import numpy as np, soundfile as sf
from kokoro import KPipeline
from nano_edge_tts import load

pipe = KPipeline(lang_code='a', model=load())   # 'a' American English, 'b' British English
audio = np.concatenate([r.audio.numpy() for r in pipe('The decoder that made this has under two million parameters.', voice='af_heart')])
sf.write('out.wav', audio, 24000)

load() returns Kokoro-82M with its decoder replaced, so it works anywhere a KModel does. Kokoro's own weights and voices are downloaded from its repository on first use. load_decoder() returns the decoder alone: decoder(asr, F0, N, style) -> waveform, the same call as Kokoro's.

Files

file	contents
nano-edge-tts.safetensors, config.json	Mapper and renderer weights (1,749,643 parameters) and their sizes
nano_edge_tts.py	The whole decoder in one file: excitation, STFTs, mapper, renderer, loader. Needs only PyTorch
paper.pdf	The paper
samples/, samples.py	The pairs above and the script that renders them
results/	Evaluation on 240 validation sentences, and whole-pipeline speed and memory

How it works

• Mapper (0.96M parameters, 80 frames/s): six ConvNeXt blocks turn Kokoro's decoder inputs into a 512-channel description of the sound.
• Renderer network (0.79M parameters, 160 frames/s): four ConvNeXt blocks output, per frame, a gain and a phase rotation for each of 301 frequency bins of a pulse train, and a gain for each of 61 bins of white noise at five sub-steps.
• Synthesiser (no parameters): a pulse train at the predicted pitch carries every harmonic, noise carries breath and consonants. The pulses use a 25 ms window, which separates harmonics; the noise uses a 5 ms window, which keeps consonant attacks sharp.

Training distils Kokoro's decoder: spectral losses against Kokoro's audio, waveform discriminators, and a critic on frozen WavLM layers that has to tell Kokoro's clips from this decoder's. None of those ship with the model.

Limitations

• Trained and evaluated on Kokoro's 28 American and British English voices. Other Kokoro languages run but are untested.
• Kokoro is the target, so this model is at best as good as Kokoro, and it inherits Kokoro's pronunciation and prosody.
• Kokoro's 28.5M-parameter text side is unchanged; it is now 94% of the parameters and 69% of the CPU time.
• A faint machine-like tone on some voiced sounds, most on two or three voices (see the paper).
• Listening results come from one listener. Speed was measured in PyTorch on a laptop CPU, not on a phone.

Licence and credit

Apache-2.0, the same as Kokoro-82M, whose text side and voices this model depends on. Kokoro-82M is by hexgrad. WavLM (Microsoft) was used only as a frozen feature extractor during training.

@misc{chumba2026nanoedgetts,
  title  = {No Network at Audio Rate: A 1.75M-Parameter Harmonic-plus-Noise Decoder for Kokoro-82M},
  author = {Mike Chumba},
  year   = {2026},
  url    = {https://huggingface.co/drmhse/nano-edge-tts}
}