Vezilka Moonshine Tiny (Macedonian, LR 5e-4)
Full fine-tune of UsefulSensors/moonshine-tiny (27M parameters) for Macedonian speech recognition, peak learning rate 5e-4.
Evaluation
Vezilka FINKI Lecture ASR Benchmark (the finki-benchmark configuration of
vezilka-ai/macedonian-speech) - five held-out FINKI
computer-science lectures with manually reviewed references; neither their audio nor
their transcripts were used in training. Normalised WER/CER in % (NFC, lower-cased,
punctuation removed); CUDA, float32, batch 1, beam 4.
| view | clips | WER | CER | S | I | D |
|---|---|---|---|---|---|---|
| full | 1,406 | 12.06 | 5.33 | 3,796 | 1,236 | 1,474 |
| speaker_balanced | 892 | 10.82 | 4.62 | 2,213 | 712 | 780 |
Usage
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="vezilka-ai/vezilka-moonshine-tiny-mk-lr5e-4")
print(asr("speech.wav", generate_kwargs={"max_new_tokens": 192})["text"])
Moonshine reads the raw 16 kHz waveform, so its cost scales with the clip length; its decoder has 194 positions, so keep generations at or below 192 tokens.
Training data
Full fine-tune for 3 epochs (1 % warm-up) on the project's training mix: the Vezilka-ASRI collection export (FINKI lectures, parliamentary sessions and NSSRM audiobooks, automatic transcripts), accepted Doniraj donations (submissions-20260831), the project's YouTube/TV transcript sets (videa_so_transkript_od_yt, alfa_audios, sitel_audios), a jargon set, and the Macedonian subsets of Mozilla Common Voice (CC0-1.0) and Google FLEURS (CC-BY-4.0).
License
Apache-2.0. The base model UsefulSensors/moonshine-tiny is released under the MIT license.
Acknowledgements
Supported by the Ministry of Education and Science of the Republic of North Macedonia (project "Utilizing AI and National Large Language Models to Advance Macedonian Language Capabilities") and co-funded by the EC/EuroHPC JU, the Ministry of Digital Transformation of the Republic of North Macedonia and the Faculty of Computer Science and Engineering (FINKI), Ss. Cyril and Methodius University in Skopje, through the VEZILKA project, Grant Agreement No. 101263128.
Citation
The data article describing this collection (Gajdov, Karamachoski, Stojcheva, Mishkovski and Mishev, Curated Macedonian speech datasets with regional transcription layers and a manually reviewed FINKI lecture benchmark) is under review; its citation will be added here on publication.