Parakeet-dk-meloa

TL;DR: a Danish fine-tune of NVIDIA's parakeet-tdt-0.6b-v3 (600M parameters, FastConformer + TDT). It writes punctuated, capitalised Danish text and cuts word errors against the base model by 27% on an independent Danish test set and by 66% on NST read speech. Trained only on data that is free for commercial use; licensed CC BY 4.0.

Usage

Install NeMo (tested with nemo_toolkit 3.0.0, PyTorch 2.14, Python 3.14):

pip install -U "nemo_toolkit[asr]"

Transcribe 16 kHz mono WAV or FLAC files:

from nemo.collections.asr.models import ASRModel

model = ASRModel.from_pretrained("kstorm77/Parakeet-dk-meloa")
output = model.transcribe(["sample_da.wav"])
print(output[0].text)

Parakeet-dk-meloa.nemo is self-contained: model weights, configuration and the tokenizer. To use a downloaded copy instead, ASRModel.restore_from("Parakeet-dk-meloa.nemo"). The model uses the base model's multilingual tokenizer unchanged, so it decodes exactly like parakeet-tdt-0.6b-v3.

Results

Word error rate (WER), greedy decoding, compared with the base model on the same test sets and the same scoring code (NeMo word_error_rate). "No punctuation/casing" lowercases text and removes punctuation before scoring, as the base model's card does; "with" scores the punctuated, capitalised output as written.

Test set	Utterances	Base (no punct./casing)	This model	Base (with)	This model (with)
Alvenir ASR eval (independent, 50 speakers)	3 888	16.40%	12.01%	26.73%	22.41%
FLEURS da_dk test	930	18.56%	16.63%	22.74%	20.93%
NST Danish test	49 174	32.22%	10.86%	38.74%	16.64%
NST test, ordinary sentences only	34 661	23.83%	7.17%	29.49%	8.06%

• The base model's own FLEURS Danish figure is 18.41%; this evaluation reproduces it at 18.56%.
• The full NST test set includes spelled-out number sequences (the model writes digits) and isolated names, which inflate its WER; the sentence-only row excludes them.
• Raw results: eval/results_nst_fleurs.json and eval/results_with_alvenir.json (the init scores there are this model).

Training

• Data (271 h after curation): NST Danish ASR database (read and dictated speech, CC0) and FLEURS da_dk train/validation (CC BY 4.0). Dictation recordings, transcripts without punctuation or casing, and characters the tokenizer cannot encode were removed; every training record whose speaker or sentence appears in a test set was removed. NST's two-microphone recordings were split into separate mono utterances.
• Recipe: full fine-tune from nvidia/parakeet-tdt-0.6b-v3, 31 000 steps (about three passes), 8-bit AdamW (bitsandbytes) at learning rate 2e-5 with 500 warmup steps and cosine decay, bf16 mixed precision, gradient clipping 1.0, Lhotse 2D duration × token bucketing sized by NeMo's OOMptimizer, frozen BatchNorm statistics. The checkpoint with the best validation WER (7.91%, step 30 000) is released.
• Hardware: one NVIDIA RTX 5070 Ti (16 GB), about 3.5 hours including validation every 1 000 steps.
• Records: training/config.yaml (complete NeMo configuration), training/run.json (provenance), training/dataset.yaml and training/dataset_report.json (dataset recipe and checks).

Limitations

• Trained on read speech only. No conversational, spontaneous or dialect test set with a commercially usable licence exists, so accuracy on such speech is not measured and is likely lower.
• Numbers come out as digits; text with spelled-out numbers scores worse.
• Inherits the base model's 16 kHz input and audio-length limits.

License and attribution

Released under CC BY 4.0, the base model's licence. When you use or redistribute it, keep the credits in ATTRIBUTION.md:

• Base model: nvidia/parakeet-tdt-0.6b-v3 by NVIDIA, CC BY 4.0. Changes: fine-tuned on Danish speech.
• NST Danish ASR Database (Nasjonalbiblioteket Språkbanken, sbr-55), CC0 1.0.
• FLEURS (Conneau et al., 2022), CC BY 4.0. Changes: audio converted to 16 kHz mono FLAC, transcripts normalised and filtered.