You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access to this model is by request and is governed by the Hebrish terms of use (https://hebrish.app/terms). Non-commercial use is free; commercial use needs a commercial licence from access-requests@hebrish.app. Please also complete the attestation form: https://hebrish.app/legal/hebrish-terms-of-use.pdf

Jachnun v1.0: MLX

MLX weights for Apple silicon of Jachnun v1.0, the Parakeet TDT 0.6B v2 fine-tune for English speech with Hebrew words dropped in ("I have to go to Bituach Leumi"), romanized, in one speaker's voice. Hebrew words come out in Latin romanization. Results, training data and limitations are on the main card. This build is Jachnun v1.0 (run r6-en30); it is replaced when the line has a new version.

Website: hebrish.app. Use of this model and every converted or quantised build of it is governed by the Hebrish terms of use. Access is by request: use the button on this page, or write to access-requests@hebrish.app. Non-commercial use (personal, research, education, open-source projects) is free; please complete the attestation form. Commercial use needs a commercial licence. Versions released under CC-BY-4.0 stay under that licence; the terms do not restrict rights it grants.

Usage

parakeet-mlx:

pip install parakeet-mlx
parakeet-mlx audio.wav --model Hebrish/Jachnun-MLX --fp32

import mlx.core as mx
from parakeet_mlx import from_pretrained
model = from_pretrained("Hebrish/Jachnun-MLX", dtype=mx.float32)
print(model.transcribe("audio.wav").text)

mlx-audio:

python -m mlx_audio.stt.generate --model Hebrish/Jachnun-MLX --audio audio.wav --output-path out --format txt

This is not an iPhone build: the Hebrish apps, and an iOS app, use the sherpa-onnx build (Hebrish/Jachnun-sherpa-onnx-int8).

Conversion and check

Converted from training run r6-en30 with the remap behind the stock mlx-community Parakeet repos (senstella's script), fp32 weights (parakeet-mlx casts to bf16 at load unless asked for fp32; it cannot load quantized weights, so there is no 8-bit build). Checked with parakeet-mlx 0.5.3 on MLX's Linux CPU backend, against the NeMo PyTorch output on held-out code-switching test clips:

	Identical to NeMo	WER vs NeMo	WER vs reference
NeMo (PyTorch)			4.9%
MLX bf16 (default)	8/10	5.4%	8.3%
MLX fp32	8/10	1.5%	3.4%

WER here is raw (cased and punctuated) on a small sample; the main card has the full evaluation. The conversion is a direct key remap (no retraining or quantization).

Use fp32. In bf16, parakeet-mlx's default, 1 of 10 check clips decoded as a run of <unk> tokens; fp32 transcribed the same clip correctly. The check ran on MLX's Linux CPU backend; Metal on a Mac is untested.