canary-1b-v2-sk

Slovak fine-tune of nvidia/canary-1b-v2 by the KInIT team. Full parameter fine-tuning on a curated Slovak speech corpus, with noise augmentation for robustness to real-world recording conditions. Listed in SLAIH, a catalog of Slovak NLP resources.

Model Details

Property	Value
Base model	nvidia/canary-1b-v2
Parameters	~978M
Architecture	FastConformer encoder + Transformer decoder (attention-based encoder-decoder), 32 encoder / 8 decoder layers
Fine-tuning method	Full fine-tuning
Language	Slovak (sk)
Task	Automatic Speech Recognition
Framework	NVIDIA NeMo
License	CC-BY-4.0

Intended Use

This model is intended for Slovak automatic speech recognition across a range of domains and recording conditions. It transcribes with punctuation and capitalization.

Out-of-scope: Non-Slovak audio, speech translation (the base model's AST capability is not retained, see Limitations), safety-critical transcription without human review.

Evaluation

Word Error Rate (WER) and Character Error Rate (CER), lower is better. Measured on two Slovak eval sets:

• CV24 - Common Voice 24.0 test split (5,239 samples, public)
• Internal - held-out KInIT set (9,317 samples, stratified by domain and speaker gender, one third clean and two thirds noise-augmented; not public)

Before vs. after fine-tuning

Model	CV24 WER ↓	CV24 CER ↓	Internal WER ↓	Internal CER ↓
kinit/canary-1b-v2-sk	8.27%	2.35%	6.37%	3.31%
nvidia/canary-1b-v2	14.00%	4.06%	14.39%	6.88%

Slovak fine-tuning reduces WER by 40.9% on Common Voice and 55.7% on the internal eval set, at the cost of multilingual and translation performance (see Limitations).

Canary vs. Parakeet

Both Slovak NeMo models were fine-tuned on the same corpus with the same noise-augmentation recipe and scored with the same pipeline, so this comparison isolates the architecture. Comparison before and after Slovak fine-tuning; WER is measured on our internal benchmark dataset (see above):

Model	Fine-tuned: WER ↓	Base: WER ↓
Canary 1B v2	6.37%	14.39%
Parakeet TDT 0.6B v3	6.62%	24.32%

Parakeet is the faster of the two.

Besides these two NVIDIA NeMo fine-tunes (Canary, Parakeet), the full KInIT ASR collection also includes six Slovak fine-tunes of Whisper, OpenAI's separate speech-recognition architecture, at sizes from tiny to large.

Training Data

Fine-tuned on an internal curated Slovak speech corpus compiled at KInIT. The corpus combines public datasets with internal KInIT recordings. Recordings containing personal data were anonymised prior to use. Samples were quality-filtered using a CER-based threshold validated against multiple ASR models. 75% of the training samples were additionally augmented with synthetic background noise for robustness to real-world recording conditions (see Training Procedure below).

Data sources
SloPalSpeech
Municipal council session recordings
Read literature
Mozilla Common Voice
TEDxSK and JumpSK Lecture Speech Corpus
FLEURS read speech
Internal KInIT recordings

Training Procedure

Hyperparameter	Value
Epochs	3
Learning rate	1e-4
LR scheduler	Polynomial decay with warmup
Optimizer	AdamW
Effective batch size	64
Precision	bf16-mixed
Gradient clipping	1.0
Noise augmentation	Applied to 75% of training samples: phone noise, background speech, background noise, white noise, and packet loss
Framework	NVIDIA NeMo with PyTorch Lightning

Training was performed on the Devana HPC cluster.

Usage

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.from_pretrained("kinit/canary-1b-v2-sk")

output = model.transcribe(
    ["audio.wav"],
    source_lang="sk",
    target_lang="sk",
    pnc="yes",
    task="asr",
)
print(output[0].text)

audio.wav must be mono — stereo input raises a channel-selector error.

License

CC-BY-4.0, inherited from the base model nvidia/canary-1b-v2 (© NVIDIA).

Limitations

• Catastrophic forgetting: Fine-tuning exclusively on Slovak ASR data significantly degrades the base model's performance on the other 24 languages it supports, as well as its speech translation (AST) capability. Use the base nvidia/canary-1b-v2 if multilingual transcription or translation is required.
• Performance may degrade on strongly accented, dialectal, or domain-specific speech not represented in the training data.
• The training data consists predominantly of short utterances. The base model's automatic chunking of long audio is inherited but was not evaluated after fine-tuning.
• Word and segment timestamps are supported by the base architecture but were not evaluated after fine-tuning.

Acknowledgements

Public datasets used in training: SloPalSpeech, Mozilla Common Voice, FLEURS, and the TEDxSK and JumpSK Lecture Speech Corpus (KEMT NLP).

(Part of the) Research results was obtained using the computational resources procured in the national project National competence centre for high performance computing (project code: 311070AKF2) funded by European Regional Development Fund, EU Structural Funds Informatization of society, Operational Program Integrated Infrastructure.