bjnortier/coreai-parakeet-tdt-0.6b-v3-float16-static

Parakeet TDT 0.6B v3 — Core AI export (float16, static)

An Apple Core AI export of NVIDIA's parakeet-tdt-0.6b-v3, for on-device speech recognition on Apple silicon.

These are not PyTorch weights. The repository holds a single archive of .aimodel assets that run through Apple's Core AI runtime on macOS 27 / iOS 27 and later. They will not load with transformers or NeMo.

Contents

parakeet-tdt-0.6b-v3_float16_static.zip (~1.16 GB) unpacks to a bundle directory:

File	Purpose
metadata.json	Bundle descriptor: architecture, vocab, decoder config
..._encoder.aimodel	FastConformer encoder
..._decoder_step.aimodel	Single-step LSTM prediction network
..._joint.aimodel	Joint network
processor/	Tokenizer

Configuration

• Architecture: Parakeet TDT (token-and-duration transducer)
• Precision: float16
• Encoder: static shape — one graph specialization is reused across inputs
• Mel bins: 128 · Subsampling: 8× · Vocab: 8193 (blank 8192)
• Durations: [0, 1, 2, 3, 4] · Max symbols/step: 10
• Audio in: 16 kHz mono float32

A static encoder is the right default: a dynamic export re-specializes for every new input length, and that cost is paid again on each process launch.

Usage

With CirceKit:

import CirceKit

let transcriber = CirceFileTranscriber(
    backend: .coreAI(.bundle(bundleURL)),
    locale: Locale(identifier: "en_US")
)
try await transcriber.prepare()
let result = try await transcriber.transcribe(fileAt: audioURL)
print(result.text)

Or directly with CoreAISpeech from apple/coreai-models:

let model = try await SpeechRecognitionModel(resourcesAt: bundleURL)
let (text, stats) = try await model.transcribe(audioURL: audioURL)

License and attribution

Released under CC-BY-4.0, the licence of the source model. The original Parakeet TDT 0.6B v3 is by NVIDIA; this repository only changes its serialization format. Refer to the upstream model card for training data, evaluation results, and intended use.