conikud-onnx

Hebrew grapheme-to-phoneme — IPA with stress, from plain unvocalized text. A single self-contained .onnx file: the wordpiece tokenizer, the chunk vocabulary and the per-letter phonetic rules all ride inside it as metadata.

pip install git+https://github.com/conikud/conikud-onnx

from conikud_onnx import G2P

g2p = G2P()  # downloads conikud_int8.onnx on first use, then cached

g2p.phonemize("קניתי ספר חדש")     # kanˈiti sˈefeʁ χadˈaʃ
g2p.phonemize("הוא ספר את הכסף")   # hˈu safˈaʁ ʔˈet hakˈesef

Hebrew omits vowels, so spelling alone rarely fixes a reading: ספר is sˈefeʁ (a book), safˈaʁ (he counted) or sapˈaʁ (a barber). The model resolves it from sentence context.

Alternatives

alternatives() keeps every word's competing readings instead of collapsing them to one answer:

g2p.alternatives("הוא ספר את הכסף", k=3)
# ספר -> safˈaʁ 0.62, safˈeʁ 0.25, sapˈaʁ 0.13

Each reading is an exact top-k beam over the chunks each letter may emit, carrying exactly one stress per word — so every variant is a well-formed pronunciation of that spelling, never an arbitrary logit sample.

Files

File	Size	Notes
conikud_int8.onnx	669 MB	int8 dynamic quantization of the MatMul weights

int8 is the right default on CPU. On GPU the quantized matmuls have no native CUDA kernels and fall back across the device boundary, losing more to memory copies than quantization saves — prefer an fp32 export there.

Accuracy

Measured on the 250-sentence internal eval set and on MILIM-Bench:

	int8	fp32
Eval WER	9.97%	9.97%
MILIM WER	16.56%	16.37%
MILIM word accuracy	83.44%	83.63%
MILIM exact sentence	71.81%	72.11%

Quantization costs at most 0.19pt of word accuracy for a 2.2× smaller file.

Lexicon and training overlap with these evaluation sets has not been audited, so these numbers do not establish a leakage-free comparison against other systems.