AshtonKem/edda-v0.2-ONNX

Edda v0.2 (ONNX)

An ONNX export of danish-foundation-models/edda-v0.2 for use with transformers.js in the browser. Edda is openai/whisper-large-v3-turbo (809M parameters, 4-layer decoder) fully fine-tuned on 2,600 hours of transcribed Danish speech; see the base model's card for training details, decoding notes, and full benchmark numbers (mean WER 8.79 across five Danish test sets on the CoRaL ASR leaderboard).

No official ONNX export existed for this model, so this conversion was done directly from the base model's safetensors checkpoint with optimum-cli export onnx (--task automatic-speech-recognition-with-past) and optimum.onnx.graph_transformations.merge_decoders.

Usage

import { pipeline } from "@huggingface/transformers";

const transcriber = await pipeline("automatic-speech-recognition", "AshtonKem/edda-v0.2-ONNX", { dtype: "q8" });
const { text } = await transcriber(audio); // 16 kHz mono Float32Array

The base model's generation_config.json already defaults to language: "da", task: "transcribe", and num_beams: 5 — the same beam search its published scores were measured with — so none of those need to be passed explicitly.

Available quantisations

dtype	size	notes
fp32	~3.5 GB	Verified: exact-match transcription against a synthesised test sentence.
q8 (dynamic int8)	~880 MB	Verified the same way. Recommended — much smaller, same output on the test case.

fp16 and q4 are not included. fp16 conversion (via onnxconverter_common) hit a genuine type-inference bug in the merged decoder's control-flow (If) node — a Add op ending up with mismatched float16/float32 operands — that persisted whether the two pre-merge decoder graphs were converted to fp16 before merging or the already-merged graph was converted after. No reliable 4-bit weight-only quantisation tool was available in the conversion environment for q4. Rather than publish something broken, both are simply absent; fp32 and q8 are what's verified to work.

License

Apache-2.0, inherited from the base model. No additional use restrictions beyond that license are attached by danish-foundation-models/edda-v0.2 itself, unlike some other models trained on CoRaL data (e.g. Harmonium/brage-v1, which adds CoRaL's Attachment A use-restrictions as an enforceable license term) — see the base model's own page for the datasets it was trained on.