Nabra-7M-Distill

Arabic text to speech in 7.48M parameters, 28.7 MB at full precision. Distilled from Nabra-82M, which is 11x larger. Small enough to run on a phone, entirely offline, with no server in the loop.

parameters	7,477,702
model file	28.7 MB (fp32)
sample rate	24 kHz
language	Modern Standard Arabic
license	Apache 2.0

Speed on CPU

Median real time factor over 24 sentences, both models on 4 CPU threads, no GPU. A phone class thread budget rather than a workstation. Text to phoneme conversion is excluded from the timing for both, since it is the same front end either way.

model	params	RTF	synth time
Nabra-7M-Distill	7.48M	0.0224	0.081 s
Nabra-82M	81.81M	0.0947	0.346 s

That is 4.2x faster than the 82M and 45x faster than realtime: one second of speech in 22 ms.

Usage

from load_model import load
model, pipeline, voice = load()
audio = next(pipeline("مَرْحَبًا بِكُم", voice=voice))[2]   # 24 kHz

Use load_model.py rather than a bare from kokoro import KModel: this config sets the decoder's hidden_channels and out_channels, which upstream Kokoro hardcodes at 1024/512, so the stock package raises a TypeError on it. The patched package is vendored in kokoro_patched/ with its defaults unchanged, so the 82M teacher still loads through it untouched.

Arabic text goes through arabic_g2p.py: normalise, optionally add tashkeel with camel-tools, then espeak-ng to IPA and clean_phonemes. Text that already carries tashkeel is used as is. The voice pack is af_msa.pt, the one the model was conditioned on during training.

Architecture

Kokoro / StyleTTS2: a 12 layer ALBERT over phonemes, a prosody predictor for duration, pitch and energy, and an ISTFTNet decoder that ends in a 20 point inverse STFT rather than more convolution.

block	params
decoder (ISTFTNet)	4,062,450
prosody predictor	2,219,572
plbert (12 shared layers)	595,520
text encoder	569,280
projection	30,880

Two details carry most of the size saving. The 12 ALBERT layers share one parameter block, so depth costs 596K rather than 7M. And duration is predicted by summing 50 sigmoid gates per phoneme instead of regressing a number.

How it was distilled

The teacher emits the per phoneme durations that generated its own audio, so the student trains against an alignment that is correct by construction. That is what makes this architecture distillable at all.

Objective: multi resolution STFT, duration L1, log mel L1, a silence term, and WavLM feature matching, with multi period and multi resolution spectrogram discriminators. The learning rate is cosine decayed to 10% of peak, which matters: at a constant rate a larger student in this family had its duration head collapse partway through while the reconstruction losses kept improving and hid it.

Limitations

Single voice, Modern Standard Arabic. Dialects are not covered. Text without tashkeel is diacritized automatically, and the model is only as good as that step, so supplying diacritized text gives better results.