---
tags:
- speech-to-text
- peft
- lora
- danish
- fine-tuned
- voxtral
- whisper
language:
- da
metrics:
- wer
- cer
base_model:
- mistralai/Voxtral-Small-24B-2507
datasets:
- CoRal-project/coral
model-index:
- name: danstral-v1
  results:
  - task:
      type: automatic-speech-recognition
      name: Automatic Speech Recognition
    dataset:
      name: CoRal read-aloud
      type: alexandrainst/coral
      split: test
      args: read_aloud
    metrics:
    - type: cer
      value: x
      name: CER
    - type: wer
      value: x
      name: WER
---

# Voxtral-Small-24B LoRA Fine-tuned on CoRaL

**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**](https://huggingface.co/mistralai/Voxtral-Small-24B-2507) with the audio encoder from [**roest-whisper-large-v1**](https://huggingface.co/CoRal-project/roest-whisper-large-v1). The decoder and audio-adapter were fine-tuned using LoRA for 2 epochs (40 hours) on the Danish [CoRaL dataset](https://huggingface.co/CoRal-project/coral), 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](https://huggingface.co/datasets/alexandrainst/coral/viewer/read_aloud/test) CER | [CoRaL](https://huggingface.co/datasets/alexandrainst/coral/viewer/read_aloud/test) WER |
|:---|---:|---:|---:|
| [hinge/danstral-v1](https://huggingface.co/hinge/danstral-v1) | 24B | **4.2% ± 0.2%** | **9.7% ± 0.3%** |
| [Alvenir/coral-1-whisper-large](https://huggingface.co/Alvenir/coral-1-whisper-large) | 1.540B | 4.3% ± 0.2% | 10.4% ± 0.3% |
| [alexandrainst/roest-315m](https://huggingface.co/alexandrainst/roest-315m) | 0.315B | 6.6% ± 0.2% | 17.0% ± 0.4% |
| [mhenrichsen/hviske-v2](https://huggingface.co/syvai/hviske-v2) | 1.540B | 4.7% ± 0.07% | 11.8% ± 0.3% |
| [openai/whisper-large-v3](https://hf.co/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](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