hviske-v6

<p align="center"> <a href="https://syv.ai" target="_blank"><img src="https://img.shields.io/badge/syv.ai-4A90E2?logoColor=white" alt="syv.ai"></a> <a href="https://platform.syv.ai" target="_blank"><img src="https://img.shields.io/badge/API-platform.syv.ai-1c1c1f" alt="API — platform.syv.ai"></a> <a href="https://github.com/Rye-A1/danish-asr-leaderboard" target="_blank"><img src="https://img.shields.io/badge/Danish%20ASR%20leaderboard-%231-4A90E2" alt="Danish ASR leaderboard"></a> <a href="https://creativecommons.org/licenses/by-nc/4.0/" target="_blank"><img src="https://img.shields.io/badge/License-CC%20BY--NC%204.0-4CAF50?logo=creativecommons&logoColor=white" alt="License CC BY-NC 4.0"></a> </p>

hviske-v6

hviske-v6 is a 1.24B-parameter Danish ASR model: a RoPE-adapted Whisper-large-v3-turbo encoder fused to a Qwen3-0.6B decoder through a frame-stacking projector. It scores a 10.19 mean WER on the Danish ASR leaderboard, the best result on the board at the time of submission.

Audio is expected at 16 kHz mono. Clips longer than 30 s are split at silences and re-joined automatically. Danish only, offline transcription (no timestamps, diarization, or streaming).

Unlike other models on the board, hviske-v6 lets you choose the output format — casing and punctuation are controllable at inference through four dedicated tokens.

Accuracy

Measured with the leaderboard harness, unmodified, on a single RTX 3090. Full test splits, no subsetting.

For comparison, the previous best on the board scores 10.68 / 5.10. FTSpeech 5.91 and Common Voice 6.21 are the best open-weights scores on the leaderboard at submission.

Throughput is 48× real-time on a 3090 at batch 4.

Usage

from processing_whisper_qwen import HviskeASR

asr = HviskeASR.from_pretrained("syvai/hviske-v6")

asr.transcribe("meeting.wav")
asr.transcribe(["a.wav", "b.mp3"], batch_size=8)

The model is custom-architecture, so trust_remote_code=True is required if you load it through AutoModel directly.

Controlling casing and punctuation

Danish ASR corpora disagree with themselves about formatting. Weighted across this model's training mix, only 8.9% of rows carry punctuation and 36.5% carry casing — FTSpeech is entirely lowercase and unpunctuated, CoRal is cased but unpunctuated, FLEURS and Common Voice are fully formatted.

Left alone, a model trained on that mix infers the format from the acoustics — it will give you lowercase, unpunctuated text for parliamentary audio no matter what you wanted. hviske-v6 makes the choice explicit instead:

asr.transcribe("clip.wav")                                  # auto — model decides
asr.transcribe("clip.wav", cased=True,  punctuated=True)    # "Han drog til Paris."
asr.transcribe("clip.wav", cased=True,  punctuated=False)   # "Han drog til Paris"
asr.transcribe("clip.wav", cased=False, punctuated=False)   # "han drog til paris"

They occupy embedding rows that were allocated but unused (the matrix is [151936, 1024]; the tokenizer stops at 151,668), so nothing was resized to add them. They are input-only — conditioning, never a prediction target. One further row is reserved the same way: <|audio_pad|> (151673) marks where audio sits in a prompt when the model is served on vLLM, and is likewise never produced.

The benchmark numbers above use the automatic path, because forcing a global style costs accuracy on domains it does not suit: cased+punc on parliamentary audio is worth +1.56 WER on FTSpeech, since that combination never occurs in training.

Serving with vLLM

trust_remote_code=True is not enough here. It loads this repo's config class; it does not run the modelling code for inference — vLLM resolves the architecture against its own registry and runs its own layers. So the repo carries a second implementation under vllm/, separate from the transformers path above and registered through vLLM's plugin entry point:

hf download syvai/hviske-v6 --local-dir hviske-v6
pip install -e hviske-v6/vllm
vllm serve syvai/hviske-v6

The hub id serves directly — no converted config and no separate model directory. The transformers path above is untouched: AutoModel still gets the implementation in this repo, because transformers only prefers a locally registered class when remote code is not trusted, which is exactly vLLM's case.

One caveat worth the two lines it takes to avoid. vLLM's own /v1/audio/transcriptions endpoint spends ~674 ms per request on audio decoding, resampling and response assembly against ~26 ms of engine work, which holds the model at about 16× realtime however many requests are in flight. Submitting prompts straight to AsyncLLM measures 281 / 370 / 445× realtime at 16 / 32 / 64 concurrent windows on one RTX 3090. vllm/README.md has the details, and a working server lives in syv-ai/transcribe.

Architecture

16 kHz → log-mel(128) → Whisper-large-v3-turbo encoder, RoPE-adapted   0.635 B
                      → stack 4 frames (one token / 80 ms)
                      → RMSNorm(5120) + Linear(5120 → 1024)            0.005 B
                      → Qwen3-0.6B-Base decoder                        0.596 B

The encoder's fixed 1500-position sinusoid table is zeroed and replaced with partial rotary embeddings over the frame index (rotary width 56 of 64 per head, θ=10000). That removes Whisper's hard 30-second window: a 4-second clip and a 27-second clip become the same problem, and the mel window can be any even length. The encoder arrives in that state as part of capacit-ai/saga-2-m.

Training

The encoder is initialised from the first release of capacit-ai/saga-2-m by Capacit A/S, published under the Apache License 2.0, and is trained further on the corpus below. The decoder is initialised from Qwen/Qwen3-0.6B-Base, and the projector is randomly initialised.

Trained on the cleaned syvai/danish-asr-unified corpus — 2,755 hours across FTSpeech, CoRal read-aloud and conversation, NST, Nota, FLEURS and Common Voice, filtered to rows whose transcripts survived a two-model agreement audit. Two full epochs, 106,689 steps on a single RTX 3090.

• Muon (encoder matrices) + AdamW (decoder, projector), warmup-stable-decay schedule
• fp32 master weights under bf16 parameters (Kahan-compensated)
• bias-corrected EMA; the released weights are the EMA
• All five benchmark test sets excluded from training by transcript hash (24,375 hashes)

Limitations

• Danish only. Other languages are not supported and will not degrade gracefully.
• Conversational speech is the weak spot (18.26 WER). Dialect drives most of it: measured across CoRal, WER ranges from 5.3 on østjysk to 23.2 on østligt sønderjysk, and older speakers score worse than younger ones. The training mix is 58% parliamentary standard Danish, which cannot teach that variation.
• Forced punctuation is weak on parliamentary audio, for the reason given above.
• No timestamps, no diarization, no streaming.

License

CC BY-NC 4.0 for this repository's own weights and card.

This model is a derivative work of capacit-ai/saga-2-m, Copyright (c) 2026 Capacit A/S, in both its inference code and its encoder weights. It was derived from a revision published under the Apache License, Version 2.0; a copy is in LICENSES/Apache-2.0.txt, and the modifications are recorded in NOTICE as section 4(b) requires.

It also derives from Qwen/Qwen3-0.6B-Base (Apache-2.0) and, through saga-2-m, from openai/whisper-large-v3-turbo (MIT). Check those repositories for their own terms before commercial use.