Moonshine-MN: Mongolian ASR
Fine-tuned UsefulSensors/moonshine-base on Mongolian (Cyrillic) speech from Mozilla Common Voice.
Performance
| Checkpoint | WER |
|---|---|
| final (step 15000) | 11.88% |
Quickstart
import torch, librosa
from transformers import MoonshineForConditionalGeneration, AutoFeatureExtractor
from mn_tokenizer import MnBPETokenizer
from huggingface_hub import hf_hub_download
model = MoonshineForConditionalGeneration.from_pretrained("orgilj/moonshine-mn").eval()
fe = AutoFeatureExtractor.from_pretrained("orgilj/moonshine-mn")
tok = MnBPETokenizer(vocab_file=hf_hub_download("orgilj/moonshine-mn", "mn_bpe.model"))
def transcribe(path, num_beams=5):
audio, _ = librosa.load(path, sr=16000)
inp = fe(audio, sampling_rate=16000, return_tensors="pt")
with torch.no_grad():
ids = model.generate(
inp.input_values,
num_beams=num_beams,
max_new_tokens=180, # under the model's max_length=194
)
return tok.decode_ids(ids[0].tolist())
if __name__ == "__main__":
print(transcribe("/workspace/data/cv-corpus-24.0-2025-12-05/mn/clips/common_voice_mn_44590402.mp3"))
Live streaming
# From finetune-moonshine-asr repo:
python scripts/stream_mn.py --model orgilj/moonshine-mn --live
Training details
- Base: moonshine-base (hidden_size=416, 8-layer enc+dec)
- Tokenizer: custom Mongolian SentencePiece BPE (vocab 2003)
- Data: Mozilla Common Voice Mongolian
- LR: 1e-4 cosine, warmup 500 steps, 15000 total steps, fp16, batch 64