TUNAR_v1_mini

Uzbek speech recognition model fine-tuned from Moonshine-tiny.

  • Output alphabet: new Uzbek Latin alphabet (ö ğ ş ç); convert with o'->ö, g'->ğ, sh->ş, ch->ç.
  • Custom 4096-token BPE tokenizer (not the original Moonshine tokenizer).
  • Input: 16 kHz mono audio, up to ~30 s per segment.
  • Validation WER: 25.30% (greedy decoding, validation subset)

Usage

import torch, soundfile as sf
from transformers import MoonshineForConditionalGeneration, AutoTokenizer, AutoFeatureExtractor

repo = "BaseLayer/TUNAR_v2_mini"
model = MoonshineForConditionalGeneration.from_pretrained(repo).eval()
tok = AutoTokenizer.from_pretrained(repo)
fe = AutoFeatureExtractor.from_pretrained(repo)

wav, sr = sf.read("audio.wav")          # 16 kHz mono
inputs = fe(wav, sampling_rate=16000, return_tensors="pt", return_attention_mask=True)
with torch.no_grad():
    ids = model.generate(**inputs, max_new_tokens=int(len(wav) / 16000 * 12) + 8)
print(tok.batch_decode(ids, skip_special_tokens=True)[0])
Downloads last month
16
Safetensors
Model size
18.8M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for BaseLayer/TUNAR_v2_mini

Finetuned
(22)
this model