You can deploy this model and start earning money today!

Noctra-labs/
parakeet-tdt-0.6b-v3-lt-coreml

Name: parakeet-tdt-0.6b-v3-lt-coreml
Availability: InStock

Views

No views yet

Lithuanian Speech-to-Text — Parakeet TDT 0.6B Fine-tuned on LIEPA (CoreML, Apple Silicon)

A CoreML build of NVIDIA's Parakeet TDT 0.6B v3, fine-tuned by Noctra Labs on the LIEPA Lithuanian speech corpus for accurate, on-device Lithuanian speech recognition on Apple Silicon Macs. This is the model Whisnap downloads for Lithuanian dictation.

This model achieves 10.87% WER on Lithuanian — a ~68% relative reduction over the stock base model (33.57% WER) — while running entirely on-device via Apple's Core ML framework with no internet connection required.

Looking for the full NeMo checkpoint? → Noctra-labs/parakeet-tdt-0.6b-v3-lt (repo id — confirm before publishing, see METADATA-NOTES.md)

Results

Model	Test Set	WER ↓
Noctra parakeet-tdt-0.6b-v3-lt (this model)	LIEPA test	10.87%
NVIDIA parakeet-tdt-0.6b-v3 (base, no fine-tune)	LIEPA test	33.57%

~68% relative WER reduction over the base model on Lithuanian speech.

Intended Use

This model is designed for:

Lithuanian speech-to-text (STT) / ASR on macOS devices with Apple Silicon (M1 and later)
On-device Lithuanian dictation and voice typing — no data leaves the device
Integration into macOS apps via Core ML, as demonstrated by Whisnap

It is not intended for:

Languages other than Lithuanian (lt)
Server-side / cloud inference (use the NeMo checkpoint for that)
Speech translation (transcription only)

Usage — Whisnap (Recommended)

The easiest way to use this model on macOS is via Whisnap — a native macOS dictation app that downloads and runs this CoreML model automatically on Apple Silicon.

Download Whisnap from whisnap.com
In Settings → Model, select Parakeet LT (Lithuanian)
Set your transcription language to Lithuanian
Press your hotkey and start speaking

Whisnap handles model download (~483 MB), CoreML compilation, and on-device inference transparently.

Usage — Direct CoreML / Swift

swift

import CoreML
import AVFoundation

// Load the compiled .mlmodelc bundle
let modelURL = Bundle.main.url(forResource: "parakeet-tdt-0.6b-v3-lt", withExtension: "mlmodelc")!
let model = try MLModel(contentsOf: modelURL)

// Pass 16 kHz mono PCM audio frames for transcription
// (Integration follows the same FluidAudio / Parakeet CoreML interface
//  used by Whisnap — see https://whisnap.com for reference implementation)

Note: The CoreML package in this repo (~483 MB) contains the encoder and decoder weights compiled for Apple Neural Engine (ANE) + Metal acceleration. It is consumed via the FluidAudio Swift wrapper that Whisnap uses. A standalone Swift inference example is planned.

Training Data

Fine-tuned on the LIEPA Lithuanian speech corpus, a high-quality, professionally recorded Lithuanian speech dataset covering read speech across multiple speakers and acoustic conditions. LIEPA is the primary benchmark corpus for Lithuanian ASR research.

Base model: nvidia/parakeet-tdt-0.6b-v3 — NVIDIA's Token-and-Duration Transducer (TDT) architecture, 0.6B parameters, originally trained on English.

Model Details

Property	Value
Architecture	Parakeet TDT (Token-and-Duration Transducer)
Base model	nvidia/parakeet-tdt-0.6b-v3
Parameters	~0.6B
File size	~483 MB (CoreML package)
Format	Apple Core ML (`.mlmodel` / `.mlmodelc`)
Target hardware	Apple Silicon (M1, M2, M3, M4 and later)
Language	Lithuanian (`lt`)
Sample rate	16 kHz mono
License	CC-BY-4.0

Limitations

Language: Lithuanian only. Performance on code-switched or heavily accented speech outside LIEPA's distribution may degrade.
Platform: CoreML format — Apple Silicon Macs only. For Linux/GPU inference use the NeMo checkpoint.
Punctuation and casing: Output follows LIEPA corpus conventions. Punctuation restoration quality depends on training data coverage.
Short utterances: Very short (< 0.5 s) or silent segments may produce empty or noisy output.
Domain: Trained primarily on read speech; spontaneous/conversational Lithuanian may have higher WER.

License

CC-BY-4.0. You are free to use, adapt, and redistribute with attribution.

Base model (nvidia/parakeet-tdt-0.6b-v3) is licensed under CC-BY-4.0. LIEPA corpus — please review liepa.lt for corpus terms.

Citation

If you use this model in research or a product, please cite:

bibtex

@misc{noctra2026parakeet_lt_coreml,
  author       = {Noctra Labs},
  title        = {Parakeet TDT 0.6B v3 Lithuanian Fine-tune — CoreML Build},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Noctra-labs/parakeet-tdt-0.6b-v3-lt-coreml}},
  note         = {Fine-tuned on the LIEPA Lithuanian speech corpus.
                  Powers on-device Lithuanian dictation in Whisnap (https://whisnap.com).}
}

Noctra-labs/parakeet-tdt-0.6b-v3-lt-coreml