parakeet-tdt-0.6b-v3-lt

Lithuanian Speech Recognition — Parakeet TDT 0.6B Fine-tuned on LIEPA (NeMo Checkpoint)

Full NeMo .nemo checkpoint of NVIDIA's Parakeet TDT 0.6B v3, fine-tuned by Noctra Labs on the LIEPA Lithuanian speech corpus. This is the research/training artifact — the primary deployment build for Apple Silicon is the CoreML package used by Whisnap.

This model achieves 10.87% WER on Lithuanian — a ~68% relative reduction from the base model's 33.57% WER — and is the first publicly released Parakeet fine-tune for Lithuanian speech-to-text.

Want to run Lithuanian ASR on a Mac without writing code? → Whisnap uses the CoreML build of this model for on-device Lithuanian dictation.

---

Results

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

---

Intended Use

This NeMo checkpoint is designed for:

• Lithuanian ASR research and experimentation using NVIDIA NeMo
• Further fine-tuning on additional Lithuanian speech data
• Server-side / GPU inference for Lithuanian speech-to-text pipelines
• Exporting to other formats (ONNX, TensorRT, CoreML via conversion)

For ready-to-use on-device Lithuanian dictation on Apple Silicon, use the CoreML build instead.

---

Usage — NeMo Inference

import nemo.collections.asr as nemo_asr

# Load directly from HuggingFace Hub
model = nemo_asr.models.ASRModel.from_pretrained(
    "Noctra-labs/parakeet-tdt-0.6b-v3-lt"
)

# Transcribe a 16 kHz mono WAV file
transcriptions = model.transcribe(["path/to/audio_lt.wav"])
print(transcriptions[0])

Requirements: nemo_toolkit[asr] (see NeMo installation).

---

Usage — NeMo Fine-tuning

import nemo.collections.asr as nemo_asr

# Load checkpoint as starting point for further fine-tuning
model = nemo_asr.models.ASRModel.from_pretrained(
    "Noctra-labs/parakeet-tdt-0.6b-v3-lt"
)

# Attach your NeMo data config and call model.fit(trainer, ...)
# See NeMo ASR fine-tuning docs for full recipe:
# https://docs.nvidia.com/nemo-framework/user-guide/latest/nemotoolkit/asr/intro.html

---

Training Data

Fine-tuned on the LIEPA Lithuanian speech corpus, a high-quality, professionally recorded Lithuanian speech dataset widely used as the standard benchmark for Lithuanian ASR research.

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

---

Model Details

---

Limitations

• Language: Lithuanian only. Performance degrades on other languages or heavily code-switched speech.
• Domain: Trained primarily on read speech (LIEPA). Spontaneous/conversational Lithuanian may yield higher WER.
• Punctuation and casing: Output follows LIEPA corpus conventions; may not generalize to all text formatting needs.
• Short utterances: Very short (< 0.5 s) or silent inputs may produce empty or noisy output.

---

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 before redistribution.

---

Citation

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

@misc{noctra2026parakeet_lt,
  author       = {Noctra Labs},
  title        = {Parakeet TDT 0.6B v3 Lithuanian Fine-tune — NeMo Checkpoint},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Noctra-labs/parakeet-tdt-0.6b-v3-lt}},
  note         = {Fine-tuned on the LIEPA Lithuanian speech corpus.
                  Powers on-device Lithuanian dictation in Whisnap (https://whisnap.com).}
}

---

Related Links

• CoreML build (Apple Silicon / Whisnap): Noctra-labs/parakeet-tdt-0.6b-v3-lt-coreml
• Whisnap — macOS Lithuanian dictation app
• Base model: nvidia/parakeet-tdt-0.6b-v3
• Noctra Labs on Hugging Face
• LIEPA Lithuanian Speech Corpus