yuriyvnv/whisper-large-v3-slovenian

Whisper-large-v3 Slovenian

Fine-tuned OpenAI Whisper-large-v3 for Slovenian automatic speech recognition, augmented with TTS-generated synthetic data.

This model is part of the paper: "Synthetic Speech Augmentation for Low-Resource Estonian and Slovenian ASR: Comparing Parakeet-TDT and Whisper" (Interspeech 2026). Paper coming soon.

Note: Fine-tuning Whisper on CommonVoice data degrades FLEURS performance for Slovenian. Zero-shot Whisper (37.02% raw / 19.18% normalized) outperforms all fine-tuned variants on FLEURS. This is because Whisper's pre-trained model handles FLEURS-style punctuation and casing, which fine-tuning on CV-style data erodes. On the in-domain CV17 test set, this model provides substantial improvements over both zero-shot and CV-only fine-tuning.

Model Description

• Architecture: Encoder-decoder Transformer (Whisper)
• Parameters: 1.55B
• Tokenizer: 51,865-token byte-level BPE
• Base model: openai/whisper-large-v3
• Fine-tuning data: CommonVoice 17.0 Slovenian + ~5,850 synthetic sentences (paraphrase + domain expansion, without morphological category)
• Training config: CV + Synth No Morph (synthetic corpus excluding morphological augmentation)

Evaluation Results

Raw WER/CER (no text normalization)

Test Set	WER	CER
CommonVoice 17 Test	15.65	3.88
CommonVoice 17 Val	15.08	3.60
FLEURS Test	40.46	11.98

Normalized WER/CER (lowercase + punctuation removal)

Test Set	WER	CER
CommonVoice 17 Test	13.24	3.48
CommonVoice 17 Val	12.98	3.24
FLEURS Test	24.52	8.32

Improvement over baselines

Comparison	CV17 Test (WER)	FLEURS Test (WER)
vs. Zero-shot	-5.55 pp	+3.44 pp*
vs. CV-only fine-tuning	-3.66 pp	-6.33 pp

*FLEURS degrades vs. zero-shot due to domain mismatch (see note above), but improves substantially vs. CV-only fine-tuning.

All CV17 improvements are statistically significant (paired bootstrap, p < 0.001, n = 100,000).

Usage

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch

# Load model and processor
model = WhisperForConditionalGeneration.from_pretrained("yuriyvnv/whisper-large-v3-slovenian")
processor = WhisperProcessor.from_pretrained("yuriyvnv/whisper-large-v3-slovenian")

# Load audio (16kHz mono)
import librosa
audio, sr = librosa.load("audio.wav", sr=16000)

# Transcribe
input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
with torch.no_grad():
    predicted_ids = model.generate(input_features, language="sl", task="transcribe")
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)

Training Details

• Optimizer: AdamW fused (lr=5e-5)
• Schedule: Linear decay with 10% warmup
• Effective batch size: 128 (64 x 2 gradient accumulation)
• Epochs: 5
• Best model: selected by eval_loss
• Precision: bf16
• Seed: 42

Synthetic Data Augmentation

The synthetic training data was generated using a three-stage pipeline:

1. Text generation: GPT-5-mini generates diverse sentences across paraphrase, domain expansion, and morphological categories
2. LLM-as-judge validation: Each sentence validated for grammaticality, naturalness, and language purity
3. Speech synthesis: OpenAI gpt-4o-mini-tts with 11-voice rotation

Dataset: yuriyvnv/synthetic_asr_et_sl

Acknowledgments

• Base model: OpenAI Whisper-large-v3
• Training data: Mozilla Common Voice 17.0
• Evaluation: Google FLEURS