Base model: openai/whisper-large-v3-turbo
Method: LoRA fine-tuning (rank 32, alpha 64, targeting q_proj/v_proj)
Training data:
- ~40,500 pseudo-label segments from a Danish phone-call transcription pipeline (self-training)
- 82 human-verified corrections collected via a manual ASR-review UI, oversampled 30× to give meaningful training weight
Epochs: 3
Final train loss: 2.73

Domain

Danish telephone calls (8 kHz narrowband, upsampled to 16 kHz). Optimised for the acoustic conditions of mobile/landline calls: background noise, codec artefacts, fast conversational speech.

Usage

python

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch

model = WhisperForConditionalGeneration.from_pretrained("x2q/whisper-large-v3-turbo-da")
processor = WhisperProcessor.from_pretrained("x2q/whisper-large-v3-turbo-da", language="danish", task="transcribe")

# audio: 16 kHz mono numpy array
inputs = processor.feature_extractor(audio, sampling_rate=16000, return_tensors="pt").input_features
with torch.no_grad():
    ids = model.generate(inputs, language="danish", task="transcribe", no_repeat_ngram_size=3)
text = processor.tokenizer.decode(ids[0], skip_special_tokens=True)

Also compatible with faster-whisper after CTranslate2 conversion.

Limitations

Optimised specifically for Danish telephone speech; may perform worse on broadcast/studio audio
Training pseudo-labels were generated by the same base model family, so systematic errors in the base model may persist
82 human-verified examples is a small correction set; further annotation would improve robustness