Danstral is a state-of-the-art 24B parameter model for Danish automatic speech recognition (ASR). It combines the decoder and audio-adapter of Voxtral-Small-24B-2507 with the audio encoder from roest-whisper-large-v1. The decoder and audio-adapter were fine-tuned using LoRA for 2 epochs (40 hours) on the Danish CoRaL dataset, using three NVIDIA L40 GPUs. While it achieves state-of-the-art performance on CoRaL, it is a massive model and likely overkill compared to Whisper-based models.

Evaluation Results

Model	Number of parameters	CoRaL CER	CoRaL WER
hinge/danstral-v1	24B	4.2% ± 0.2%	9.7% ± 0.3%
Alvenir/coral-1-whisper-large	1.540B	4.3% ± 0.2%	10.4% ± 0.3%
nvidia/parakeet-rnnt-110m-da-dk	0.110B	-	10.7%
alexandrainst/roest-315m	0.315B	6.6% ± 0.2%	17.0% ± 0.4%
mhenrichsen/hviske-v2	1.540B	4.7% ± 0.07%	11.8% ± 0.3%
openai/whisper-large-v3	1.540B	11.4% ± 0.3%	28.3% ± 0.6%

Limitations

Danstral-v1 is huge. It's 16x the size of coral-1-whisper-large with only modest performance improvements. However, the LoRA adapter itself is only 25 million parameters.
Danstral-v1 is a fine-tuned version of voxtral-small-24b, whose encoder is a fine-tuned version of mistral-small-24b. Mistral does not disclose its training datasets, but it is likely that Danish Wikipedia articles were used. Since the CoRaL test split also contains read-aloud samples from Danish Wikipedia, there is a risk of data leakage, which could influence the test scores.
The model was fine-tuned solely on the CoRaL v1 dataset, so performance may deteriorate for other data sources.

Future Work and Ideas

Further optimization. The state-of-the-art performance was achieved with a 25M parameter LoRA adapter. I only conducted a few experiments, and there are likely more performance gains to be had by tweaking the LoRA configuration or by conducting a full parameter fine-tune.
Knowledge distillation. Danstral-v1 can be used for knowledge distillation to train smaller models.

How to Use

See https://github.com/ChristianHinge/danstral for the training script.

python

from transformers import VoxtralForConditionalGeneration, AutoProcessor, WhisperForConditionalGeneration
import torch
from peft import PeftModel
from datasets import load_dataset, Audio

repo_id = "mistralai/Voxtral-Small-24B-2507"

processor = AutoProcessor.from_pretrained(repo_id)
model = VoxtralForConditionalGeneration.from_pretrained(repo_id, torch_dtype=torch.bfloat16, device_map="auto",attn_implementation="flash_attention_2")

# Load audio encoder
whisper_model = WhisperForConditionalGeneration.from_pretrained(
    "CoRal-project/roest-whisper-large-v1",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2"
)

whisper_encoder_state_dict = whisper_model.model.encoder.state_dict()
model.audio_tower.load_state_dict(whisper_encoder_state_dict)

# Load LoRA adapters
model = PeftModel.from_pretrained(model, "hinge/danstral-v1")

coral = load_dataset("CoRal-project/coral", "read_aloud")
coral = coral.cast_column("audio", Audio(sampling_rate=16000))

for i in range(10):
    sample = coral["test"][i]
    audio_data = sample['audio']
    ground_truth = sample['text']

    inputs = processor.apply_transcription_request(language="da", audio=audio_data['array'], format=["WAV"], model_id=repo_id)
    inputs = inputs.to("cuda:0", dtype=torch.bfloat16)

    outputs = model.generate(**inputs, max_new_tokens=256,do_sample=False)
    decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)

    print(f"Ground Truth: {ground_truth}")
    print(f"Prediction: {decoded_outputs[0]}")
    print("-" * 40)

Shoutouts

Viktor Stenby Johansson and Rasmus Asgaard for ASR hackathon and ideation
The CoRal project and Alexandra Institute for curating Danish datasets and leading the effort in Danish NLP