Lapillus-v1

Lapillus-v1 is an Armenian speech recognition model created by David Ghazaryan.

It is fine-tuned from nvidia/stt_hy_fastconformer_hybrid_large_pc (FastConformer, hybrid RNN-T + CTC, 115M parameters) on a private dataset of Eastern Armenian speech. It outputs text with punctuation and capitalisation.

Results

ArmBench-ASR, evaluated by Metric AI. Normalised WER / CER (%), lower is better.

Dataset NVIDIA base WER Lapillus-v1 WER NVIDIA base CER Lapillus-v1 CER
Combined 10.23 9.03 3.06 2.76
MCV26 5.22 5.02 1.17 1.11
FLEURS 7.29 6.46 2.33 1.93
Poems 11.68 10.09 3.22 3.02
Movies 37.42 33.05 18.54 17.32
Infocom May 2026 9.22 6.92 2.50 1.88

Usage

# pip install "nemo_toolkit[asr]"
import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.from_pretrained("davitlive5/Lapillus-v1")
print(model.transcribe(["audio.wav"])[0].text)  # 16 kHz mono

For faster CTC decoding: model.change_decoding_strategy(decoder_type="ctc"). Split audio longer than about 30 seconds into shorter segments before transcribing.

Training

Fine-tuned for 8 epochs on a private Eastern Armenian speech dataset with noise, codec and speed augmentation. The final model is an average of the 5 best checkpoints.

Limitations

  • Eastern Armenian only. Expect more errors on Western Armenian.
  • Accuracy drops on film dialogue, heavy background noise, rare names and specialised terms.

License

CC-BY-4.0, derived from NVIDIA's stt_hy_fastconformer_hybrid_large_pc (CC-BY-4.0).

Citation

@misc{ghazaryan2026lapillus,
  title        = {Lapillus-v1: Armenian Speech Recognition},
  author       = {Ghazaryan, David},
  year         = {2026},
  url          = {https://huggingface.co/davitlive5/Lapillus-v1}
}
Downloads last month
35
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for davitlive5/Lapillus-v1

Finetuned
(2)
this model
Finetunes
1 model

Evaluation results