Milo-ASR: Dansk ASR Model

Milo-ASR er en dansk speech-to-text model baseret på Qwen3-ASR-1.7B, finetuned til at forstå dansk - både oplæst tale og samtaler/podcasts.

Modellen er trænet på CoRal v2 + danske podcast-data, så den klarer sig godt på tværs af domæner. De fleste andre modeller er kun gode til enten det ene eller det andet.

Resultater

CoRal v2 (oplæst tale, 10.370 samples)

Model	WER	CER
hviske-v2 (Whisper v2)	17.40%	7.96%
hviske-v3 (Whisper v3)	21.62%	9.22%
Milo-ASR	23.24%	11.17%
Whisper v3 Turbo	40.35%	15.51%
Qwen3-ASR base	46.28%	19.78%

Podcast (samtaler, 500 samples)

Model	WER	CER
Milo-ASR	21.82%	15.64%
hviske-v2 (Whisper v2)	50.67%	38.31%
Whisper v3 Turbo	67.03%	45.98%
Qwen3-ASR base	67.52%	47.71%
hviske-v3 (Whisper v3)	67.65%	50.12%

Milo-ASR er den eneste model der klarer begge domæner godt. På podcasts er den 2.3x bedre end næstbedste model (hviske-v2).

Plots

---

Quick Start

pip install qwen-asr transformers torch

from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "pluttodk/Milo-ASR",
    dtype="bfloat16",
    device_map="cuda:0",
)

results = model.transcribe(
    audio="path/to/danish_audio.wav",
    language="Danish",
)

print(results[0].text)

Batch Transcription

audio_files = ["audio1.wav", "audio2.wav", "audio3.wav"]
results = model.transcribe(audio=audio_files, language="Danish")

for r in results:
    print(r.text)

Timestamps

model = Qwen3ASRModel.from_pretrained(
    "pluttodk/Milo-ASR",
    forced_aligner="Qwen/Qwen3-ForcedAligner-0.6B",
    dtype="bfloat16",
    device_map="cuda:0",
)

results = model.transcribe(
    audio="path/to/audio.wav",
    language="Danish",
    return_time_stamps=True,
)

for item in results[0].time_stamps.items:
    print(f"{item.start_time:.2f}s - {item.end_time:.2f}s: {item.text}")

Streaming (vLLM)

model = Qwen3ASRModel.LLM(
    model="pluttodk/Milo-ASR",
    gpu_memory_utilization=0.8,
)

state = model.init_streaming_state(language="Danish", chunk_size_sec=2.0)

for audio_chunk in audio_stream():
    state = model.streaming_transcribe(audio_chunk, state)
    print(state.text)

state = model.finish_streaming_transcribe(state)

---

Træningsdetaljer

Modellen er finetuned i to stages:

1. Stage 1: Qwen3-ASR-1.7B finetuned på CoRal v2 (~250K samples, 3 epochs, lr=2e-5)
2. Stage 2: Fortsat fra stage 1 checkpoint på podcast + Azure podcast data (~141K samples, 8 epochs, lr=1e-5, cosine schedule)

Parameter	Stage 2
Learning rate	1e-5
Batch size	8 (x4 grad acc = 32 effective)
Epochs	8
LR scheduler	Cosine
Warmup ratio	0.1
Weight decay	0.01
Precision	bfloat16
Training steps	35,560

---

Ting nedarvet fra Qwen3-ASR

• Streaming/real-time via vLLM
• Sang detection (baggrundsmusik)
• Word-level timestamps
• 30+ sprog (dansk optimeret)
• Op til 20 min audio pr. request

---

Citation

@misc{Milo-ASR,
  author = {Rønnelund, Mathias Oliver Valdbjørn},
  title = {Milo-ASR: Danish ASR Model based on Qwen3-ASR},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/pluttodk/Milo-ASR}
}

Acknowledgements

• Qwen Team for Qwen3-ASR
• Alexandra Institute for CoRal v2