drmhse/compressed-kokoro

Compressed Kokoro

Smaller drop-in decoders for Kokoro-82M. Kokoro's text side (ALBERT, text encoder, prosody predictor) and all 28 English voices stay exactly as they are, so pronunciation, timing and prosody are identical to Kokoro's. Only the waveform decoder, which holds 65% of Kokoro's parameters, is replaced. Each decoder was pruned from Kokoro's own and then distilled from it.

variant	whole model	decoder speed¹	voiced artifact rate²	in short
Kokoro-82M	81.8M (100%)	1.00×	—	reference
H512_G512	59.3M (73%)	1.03×	3.8%	Sounds like Kokoro: a blind listener could not tell them apart (18 of 20 pairs "same").
H512_G256_S128-128	46.8M (57%)	1.39×	5.4%	Balanced: much smaller and faster, close to Kokoro.
H512_G256	44.0M (54%)	2.65×	18.2%	Fastest. A faint hoarse whistle on some voiced sounds.

¹ Decoder real-time factor relative to Kokoro's decoder on the same input. Apple M4 GPU, PyTorch MPS, batch 1, median over interleaved runs. ² Share of voiced frames with a narrowband 2–3 kHz excess of more than 15 dB over Kokoro: the paired measure of the "hoarse whistle" artifact (validation set, lower is better).

Quick start

pip install kokoro==0.9.4 soundfile huggingface_hub
hf download drmhse/compressed-kokoro --local-dir compressed-kokoro --exclude "samples/*"
cd compressed-kokoro

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

model = load('H512_G512')       # or 'H512_G256_S128-128', 'H512_G256'
pipe = KPipeline(lang_code='a', model=model, repo_id='hexgrad/Kokoro-82M')   # 'a' = US English, 'b' = UK English
audio = np.concatenate([a.numpy() for _, _, a in pipe('Hello from a smaller Kokoro.', voice='af_heart')])
sf.write('hello.wav', audio, 24000)

slim.load builds Kokoro's KModel from the official checkpoint and swaps in the slim decoder, so the rest of the kokoro package (voices, KPipeline, speed=) works unchanged. Each variant folder holds decoder.safetensors (f32, weight norm folded into plain weights) and config.json (widths and training steps).

Results

Each slim decoder is compared with Kokoro's own decoder on identical inputs: the same phonemes, durations, prosody and excitation.

Test set: 112 sentences from three public-domain books not used in training, 4 per voice for all 28 voices (13.9 min).

	Kokoro	H512_G512	H512_G256_S128-128	H512_G256
Whisper small.en WER	2.51%	2.39%	2.27%	2.43%
log-mel L1 vs Kokoro ↓	—	0.179	0.245	0.274
waveform SNR vs Kokoro ↑	—	8.7 dB	7.3 dB	6.2 dB
long-term spectrum cosine (mean / worst) ↑	—	0.9983 / 0.992	0.9969 / 0.990	0.9954 / 0.982
voiced artifact rate ↓	—	4.8%	6.9%	20.7%

No variant loses intelligibility. The WER differences (56–62 errors) are within the noise of a test this size.

Blind listening test: Kokoro vs H512_G512. 22 pairs, one sentence per voice, random order, loudness-matched, including two hidden catch pairs (Kokoro against itself). The listener answered "same" on 18 of the 20 real pairs, preferred Kokoro once and the slim decoder once, and answered "same" on both catch pairs.

Per-file reports are in results/; listen in samples/.

How it was made

Kokoro's decoder has two parts: a stack of AdaIN residual blocks (width 1024) and an iSTFTNet waveform generator (two upsampling stages, 256 and 128 channels, then an inverse STFT). The variant names give the widths: H is the AdaIN width, G the generator input, and S the per-stage generator widths when they differ from the default G/2, G/4.

1. Pruning initialisation. Channels are ranked by weight importance and kept consistently across every tensor that reads or writes them, so each student starts from Kokoro's own strongest channels.
2. Shared-excitation distillation. The generator is driven by a random sine-and-noise excitation. During training Kokoro and the student receive the same excitation draw on the same 1.6 s window, which makes Kokoro's waveform an exact, learnable target, phase included.
3. Losses: multi-resolution STFT and log-mel, plus a multi-resolution complex STFT term. The shared excitation is what makes that term meaningful, and it removed a crackling artifact without any adversarial (GAN) training.

Data: 21.5 hours of Kokoro's own output: 12,000 sentences from 18 public-domain books, all 28 voices, 10% at varied speed. One Apple M4 laptop GPU; 5,000 training steps (about 1.7 h) for the two wide-generator variants and 30,000 for H512_G256. Code and recipe are in training/.

What we learned about compressing this decoder

• The AdaIN blocks compress well. Halving them (1024 → 512) costs nothing audible. They set the model's size but barely its speed.
• The generator's last stage decides both speed and quality. It runs at 4,800 frames per second and does most of the work. At 64 channels instead of Kokoro's 128, a voiced hoarse whistle appears that more training never removed. Restoring that single stage (+2.9M parameters) brings the artifact rate from 18.2% to 5.4%, and the speed-up from 2.65× to 1.39×.
• Mel distance can mislead. Between two of our models it disagreed with the listener and with every other measure; a phase-aware loss looked slightly worse by mel and sounded "better by far".

Limitations

• One listener, 20 pairs per comparison, no MUSHRA-style anchors. Treat "indistinguishable" as strong evidence, not proof. H512_G256_S128-128 has not yet been through a blind test.
• Speed-ups are for the decoder in PyTorch on one machine. Whole-pipeline gains are smaller because the text side is unchanged.
• English only (Kokoro's American and British voices). Quality is capped by Kokoro, including its mispronunciations.

Files

path	contents
slim.py	Loader: Kokoro's KModel with a slim decoder.
<variant>/decoder.safetensors, <variant>/config.json	Decoder weights and widths.
samples/	Kokoro and each variant on six test-set sentences.
results/	Evaluation and speed reports (JSON) behind every number above.
training/	Data preparation, teacher cache, distillation, evaluation, perceptual judge and benchmark scripts.

License and attribution

Apache-2.0, like Kokoro-82M. The decoders are derived from Kokoro-82M by hexgrad (Apache-2.0), which builds on StyleTTS 2 and iSTFTNet. Training text is from Project Gutenberg (public domain).