oddadmix/Nabra-7M-Distill-ONNX

Nabra-7M-Distill (ONNX)

ONNX export of oddadmix/Nabra-7M-Distill, a 7.48M-parameter Arabic (MSA) text-to-speech model. These files are used by the nabra-tts npm package, which runs the model fully client-side with ONNX Runtime Web (WebGPU / WASM).

npm i nabra-tts

import { NabraTTS } from 'nabra-tts';
const tts = await NabraTTS.load();
const { audio, sampleRate } = await tts.generate('مَرْحَبًا بِكُمْ'); // 24 kHz mono

Files

file	size
onnx/model.onnx	30.5 MB	fp32
onnx/model_fp16.onnx	16.0 MB	fp16 weights, fp32 inputs/outputs
voices/af_msa.bin	0.5 MB	voice pack af_msa, raw little-endian float32 [510, 256]
config.json		original config (vocab)

Inputs / outputs

name	type	shape
input_ids	int64	[1, T]	phoneme token ids, wrapped in 0 … 0; T ≤ 512
style	float32	[1, 256]	row len(phonemes) - 1 of the voice pack
speed	float32	[1]	1.0 = normal
waveform	float32	[N]	24 kHz audio
durations	int64	[T]	predicted frames per token

The phonemes come from espeak-ng (ar) followed by clean_phonemes from arabic_g2p.py in the source repository. ʕ and ħ map to token ids 7 and 8.

Validation

The export is checked against PyTorch on held-out sentences. The ISTFTNet source injects random noise, so samples never match bit-for-bit. The comparison uses mean log-spectrogram distance, with a second PyTorch run giving the noise floor:

	mean Δ log-spec
PyTorch rerun (noise floor)	0.350
model.onnx	0.350
model_fp16.onnx	0.360

Output lengths (durations) are identical. Dynamic int8 quantization was tested and dropped. It saved only 27% and audibly degraded the output.

Export script: export/export_onnx.py in the nabra-tts repository.