parakeet-rnnt-120m-eou-coreml-int8

CoreML conversion of nvidia/parakeet_realtime_eou_120m-v1 — INT8 PER CHANNEL SYMMETRIC quantized.

Architecture	RNNT
Language	English
Sample rate	16000 Hz
Max audio	15.0s
Vocab size	1026
Framework	NVIDIA NeMo → CoreML (coremltools)

Components

File	Component	Best compute
parakeet_mel_encoder.mlpackage	mel_encoder	ANE / GPU
parakeet_decoder.mlpackage	decoder	CPU only
parakeet_joint_decision_single_step.mlpackage	joint_decision_single_step	ANE / GPU

Usage

pip install ovos-stt-plugin-coreml

from ovos_stt_plugin_coreml import CoremlSTT
from ovos_plugin_manager.utils.audio import AudioFile

stt = CoremlSTT(config={"metadata": "metadata.json"})

with AudioFile("speech.wav") as f:
    audio = f.read()
print(stt.execute(audio))

Source model

nvidia/parakeet_realtime_eou_120m-v1