syvai/hviske-v5

Hviske v5 — Danish ASR

Fine-tuned CohereLabs/cohere-transcribe-03-2026 (2B parameters) for Danish speech recognition.

Trained on 3.5M samples (16,000 hours) of Danish speech across 7 datasets.

Training progression

Version	Training data	Avg WER
Base model	—	>100% (no Danish pretraining)
hviske-v1	CoRal-v3	~20%
hviske-v4	+ nota + ftspeech	6.7% (parliamentary)
hviske-v5	+ VoxPopuli + nst-da + Common Voice	14.1% (multi-domain avg)

Evaluation results

Evaluated on 200 samples per dataset (WER / CER):

Dataset	Domain	WER	CER
VoxPopuli	European Parliament	11.8%	6.3%
nota	Broadcast media	5.7%	1.8%
ftspeech	Danish Parliament	6.4%	3.1%
CoRal-v3 read_aloud	Read-aloud speech	19.2%	7.0%
CoRal-v3 conversation	Conversational	17.1%	9.3%
nst-da	General Danish	14.0%	8.4%
Common Voice 17	Crowd-sourced	24.3%	7.4%
Average		14.1%	6.2%

Training details

• Architecture: Full fine-tuning (encoder + decoder), encoder BatchNorm in eval mode
• Optimizer: AdamW 8-bit, LR 2.5e-5, cosine schedule with 500-step warmup
• Batch size: 32 (4 × 8 gradient accumulation)
• Data mixing: Shuffle buffer (20k samples) across all datasets for balanced domain exposure
• Audio: 16kHz mono, max 20s per sample
• Hardware: Single NVIDIA RTX 3090 (24GB)

Training data

Dataset	Samples	Hours	Description
VoxPopuli Danish	1,775,578	~13,600	European Parliament recordings
ftspeech	995,677	~1,400	Danish Parliament (Folketinget)
CoRal-v3 read_aloud	299,255	~400	Read-aloud Danish speech
nst-da	182,605	~250	NST Danish speech corpus
CoRal-v3 conversation	147,249	~200	Conversational Danish
nota	98,600	~270	Danish broadcast media
Common Voice 17	3,484	~5	Crowd-sourced Danish
Total	~3.5M	~16,000

A unified version of all training data is available at syvai/danish-asr-unified.

Usage

Installation

pip install transformers torch soundfile librosa

Basic transcription

from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
import soundfile as sf

# Load model
processor = AutoProcessor.from_pretrained("syvai/hviske-v5", trust_remote_code=True)
model = AutoModelForSpeechSeq2Seq.from_pretrained("syvai/hviske-v5", trust_remote_code=True)
model = model.to("cuda")  # optional, for GPU inference

# Load audio (must be 16kHz mono)
audio, sr = sf.read("audio.wav")

# Transcribe
transcriptions = model.transcribe(
    processor=processor,
    audio_arrays=[audio],
    sample_rates=[sr],
    language="da",
    punctuation=True,
)
print(transcriptions[0])

Batch transcription

import soundfile as sf

files = ["file1.wav", "file2.wav", "file3.wav"]
arrays, rates = [], []
for f in files:
    audio, sr = sf.read(f)
    arrays.append(audio)
    rates.append(sr)

transcriptions = model.transcribe(
    processor=processor,
    audio_arrays=arrays,
    sample_rates=rates,
    language="da",
    punctuation=True,
)
for f, t in zip(files, transcriptions):
    print(f"{f}: {t}")

High-throughput inference with vLLM

For production workloads, serve the model with vLLM for significantly higher throughput:

pip install vllm
vllm serve syvai/hviske-v5 --trust-remote-code

Then send requests:

import requests, base64, soundfile as sf, io

audio, sr = sf.read("audio.wav")
buf = io.BytesIO()
sf.write(buf, audio, sr, format="WAV")
audio_b64 = base64.b64encode(buf.getvalue()).decode()

response = requests.post("http://localhost:8000/v1/chat/completions", json={
    "model": "syvai/hviske-v5",
    "messages": [{"role": "user", "content": [
        {"type": "input_audio", "input_audio": {"data": audio_b64, "format": "wav"}}
    ]}],
})
print(response.json()["choices"][0]["message"]["content"])