EverParakeet v1.1, full precision

The full-precision (fp32) fine-tuned weights of Everpaper's EverParakeet v1.1, as a NeMo .nemo file and as Hugging Face transformers files. This is the source the other EverParakeet builds (Core AI, Core ML and MLX) are made from. It is based on NVIDIA's parakeet-tdt-0.6b-v3; all credit for the architecture, tokenizer and multilingual training goes to NVIDIA.

The EverParakeet family

Repository	What it is	Runs on
everparakeet-tdt-0.6b-v3-coreai	Standard: 4-bit encoder, Core AI and Core ML	Neural Engine
everparakeet-tdt-0.6b-v3-compact	Compact: smaller, mixed 2/4-bit encoder, most accurate on FLEURS-25	Neural Engine
everparakeet-tdt-0.6b-v3-mini	Mini (experimental): smallest, mixed 2/4-bit encoder	Neural Engine
everparakeet-tdt-0.6b-v3-mlx	Fine-tuned weights for parakeet-mlx (bf16)	GPU through MLX
everparakeet-tdt-0.6b-v3	Full precision: .nemo and Hugging Face transformers	any

The generation (currently 1.1) is in the git tags, not in the repository names.

Load it

NeMo

pip install -U "nemo_toolkit[asr]" huggingface_hub
hf download EverPaper/everparakeet-tdt-0.6b-v3 everparakeet-tdt-0.6b-v3.nemo --local-dir .

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.restore_from("everparakeet-tdt-0.6b-v3.nemo")
print(model.transcribe(["audio.wav"])[0].text)

Hugging Face transformers

Requires a transformers release that includes the Parakeet TDT model class (the files were exported with transformers 5.6; the loader check used 5.9).

from transformers import AutoModelForTDT, AutoProcessor

repo = "EverPaper/everparakeet-tdt-0.6b-v3"
processor = AutoProcessor.from_pretrained(repo)
model = AutoModelForTDT.from_pretrained(repo)

Audio input follows the original model's processor (16 kHz mono). When decoding with batch_decode, remove the <blank> markers from the text.

model.safetensors has exactly the same tensor names, shapes and dtypes as NVIDIA's file (723 tensors); 651 of them differ from the original (fine-tuned) and 72 are identical to it.

Accuracy

How to read these numbers. Every EverParakeet figure below was measured by us on one Mac with the same test harness and the same text normaliser. Lower WER (word error rate) is better. The fine-tuning data included the FLEURS training split, so part of the FLEURS gain is in-domain and will not carry over fully to other audio. The NVIDIA column is the number NVIDIA publishes with its own text normalisation, so it is not comparable one to one with ours. This card makes no claim to beat the original model.

This repository holds the weights before compression. We measured the bf16 MLX conversion of these weights on our harness: 12.78 mean WER on FLEURS-25 (the MLX original control scores 13.23) and 5.56 on the English Open ASR Leaderboard nine-set mean (control 5.60). We did not run a separate fp32 NeMo or transformers pass over the full benchmark sets. The compressed Neural Engine builds are slightly less accurate on FLEURS-25 than the bf16 weights (Standard 13.39, Compact 12.85, Mini 14.10), and see their cards for details.

FLEURS test, 25-language mean WER	%
EverParakeet v1.0	13.75
EverParakeet Standard v1.1	13.39
EverParakeet Compact	12.85
EverParakeet Mini (experimental)	14.10
FluidAudio Parakeet v3 Core ML	13.59
EverParakeet v1.1 MLX (bf16)	12.78 (MLX original control 13.23)
NVIDIA published	11.97 (NVIDIA's own normalisation, not comparable one to one)

License and credit

The model weights, tokenizer and other adapted model material are derived from NVIDIA's Parakeet TDT 0.6B v3 and are released under CC-BY-4.0, as the original. All credit for the architecture, the tokenizer and the multilingual training goes to NVIDIA. This is not an official NVIDIA distribution and NVIDIA does not endorse it. See NOTICE for the list of modifications. Three Swift files adapted from Apple's Core AI speech sample code remain under Apple's BSD 3-clause licence (swift/THIRD_PARTY_NOTICES).

See NOTICE for the modifications (fine-tuning and the re-export of the weights).