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.

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 Estonian + ~5,850 synthetic sentences (LLM-generated text + OpenAI TTS)
Training config: CV + Synth All (full synthetic corpus with quality filtering)

Evaluation Results

Raw WER/CER (no text normalization)

Test Set	WER	CER
CommonVoice 17 Test	26.46	6.24
CommonVoice 17 Val	25.25	5.63
FLEURS Test	36.56	7.07

Normalized WER/CER (lowercase + punctuation removal)

Test Set	WER	CER
CommonVoice 17 Test	24.00	5.69
CommonVoice 17 Val	23.11	5.15
FLEURS Test	13.89	3.20

Improvement over baselines

Comparison	CV17 Test (WER)	FLEURS Test (WER)
vs. Zero-shot	-7.94 pp	-4.16 pp
vs. CV-only fine-tuning	-2.92 pp	-1.95 pp

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

Usage

python

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch

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

# 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="et", 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:

Text generation: GPT-5-mini generates diverse sentences across paraphrase, domain expansion, and morphological categories
LLM-as-judge validation: Each sentence validated for grammaticality, naturalness, and language purity
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