varsan-g/hviske-v2-onnx

Hviske v2 ONNX

ONNX export of syvai/hviske-v2 for use with Transformers.js in the browser.

Hviske v2 is a state-of-the-art Danish speech-to-text model, fine-tuned from openai/whisper-large-v3 on CoRal and Common Voice 17.0 datasets. Developed by syv.ai.

Performance

Model	Parameters	CoRal CER	CoRal WER
syvai/hviske-v2	1540M	4.7%	11.8%
alexandrainst/roest-315	315M	6.6%	17.0%
openai/whisper-large-v3	1540M	11.4%	28.3%

Usage with Transformers.js

import { pipeline } from '@huggingface/transformers'

const transcriber = await pipeline(
  'automatic-speech-recognition',
  'varsan-g/hviske-v2-onnx',
)

const result = await transcriber(audioData)
console.log(result.text)

Model Details

• Base model: openai/whisper-large-v3
• Fine-tuned by: syv.ai
• Language: Danish (da)
• License: OpenRAIL
• Format: ONNX (int8 quantized, single-file, no external data)
• Encoder: 615 MB
• Decoder: 932 MB

Conversion

Exported from PyTorch using Optimum and quantized with ONNX Runtime dynamic int8 quantization for browser compatibility:

optimum-cli export onnx --model syvai/hviske-v2 --task automatic-speech-recognition output/

File Structure

onnx/
  encoder_model.onnx           (615 MB)
  decoder_model_merged.onnx    (932 MB)