Modelpublic

FredrikKarlssonSpeech/wav2vec2-large-voxrex-swedish-onnx

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes26downloads

Model Card

wav2vec2-large-voxrex-swedish — ONNX

ONNX export of KBLab/wav2vec2-large-voxrex-swedish, a 317M-parameter Wav2Vec2ForCTC model fine-tuned for Swedish automatic speech recognition using the VoxRex corpus.

Exported with 🤗 Optimum (v2.1.0), opset 17. FP16/INT8/UINT8/Q4 quantized variants included.

Available Variants

File	Size	Notes
`onnx/model.onnx`	1262 MB	FP32 — full precision baseline
`onnx/model_fp16.onnx`	632 MB	FP16 internal weights, float32 I/O — recommended
`onnx/model_int8.onnx`	317 MB	INT8 dynamic — validate on Swedish audio before use
`onnx/model_uint8.onnx`	317 MB	UINT8 dynamic — closest to FP32 speed on CPU
`onnx/model_q4.onnx`	241 MB	Q4 MatMul — most aggressive compression

FP16 is the recommended quantized variant. Dynamic INT8/UINT8/Q4 show logit drift without Swedish calibration data; FP16 produces text-identical output to FP32.

Usage

With 🤗 Optimum (recommended)

python

from optimum.onnxruntime import ORTModelForCTC
from transformers import Wav2Vec2Processor
import soundfile as sf
import torch

model_id = "FredrikKarlssonSpeech/wav2vec2-large-voxrex-swedish-onnx"

processor = Wav2Vec2Processor.from_pretrained(model_id)
model = ORTModelForCTC.from_pretrained(model_id, subfolder="onnx", file_name="model.onnx")

audio, sr = sf.read("your_audio.wav")
assert sr == 16000, "Resample to 16kHz first"

inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits

predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)[0]
print(transcription)

With ONNX Runtime directly

python

import onnxruntime as ort
import numpy as np
import soundfile as sf
import torch
from transformers import Wav2Vec2Processor
from huggingface_hub import hf_hub_download

model_id = "FredrikKarlssonSpeech/wav2vec2-large-voxrex-swedish-onnx"
processor = Wav2Vec2Processor.from_pretrained(model_id)

# Download and load ONNX model (use model_fp16.onnx for half the size)
model_path = hf_hub_download(model_id, filename="onnx/model.onnx")
sess = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])

audio, sr = sf.read("your_audio.wav")
inputs = processor(audio, sampling_rate=16000, return_tensors="np")

logits = sess.run(["logits"], {"input_values": inputs["input_values"]})[0]
predicted_ids = np.argmax(logits, axis=-1)
transcription = processor.batch_decode(torch.tensor(predicted_ids))[0]
print(transcription)

Using FP16 variant (half the size, same accuracy)

python

from optimum.onnxruntime import ORTModelForCTC
from transformers import Wav2Vec2Processor

model_id = "FredrikKarlssonSpeech/wav2vec2-large-voxrex-swedish-onnx"
processor = Wav2Vec2Processor.from_pretrained(model_id)

# FP16: 632 MB, text-identical to FP32
model = ORTModelForCTC.from_pretrained(model_id, subfolder="onnx", file_name="model_fp16.onnx")

Resampling audio to 16kHz

python

import torchaudio

waveform, sample_rate = torchaudio.load("your_audio.wav")
if sample_rate != 16000:
    waveform = torchaudio.functional.resample(waveform, sample_rate, 16000)
audio = waveform.squeeze().numpy()

Performance

Benchmarked on M1 Mac (Apple Silicon), 10 seconds of 16kHz audio, 5 runs after 2 warmup:

Backend	Inference time	Speedup
PyTorch CPU (Accelerate/vecLib)	0.67s	1.00x
ONNX FP32 — CPU EP	1.12s	0.60x
ONNX UINT8 — CPU EP	0.71s	0.94x
ONNX FP16 — CPU EP	1.15s	0.58x
ONNX FP32 — CoreML EP	5.07s	0.13x

On Apple Silicon, PyTorch uses Apple's Accelerate framework (vecLib BLAS) which outperforms ONNX Runtime's generic CPU path. On x86/Linux, ONNX Runtime with INT8 typically achieves 2–4x speedup over PyTorch.

Accuracy

FP32 ONNX output is numerically identical to PyTorch (max logit diff: 0.002, mean: 0.00006, text match: exact).

Variant	Text match vs FP32	Max logit diff
FP16	✓ Identical	1.31
INT8 dynamic	✗ Differs	8.56
UINT8 dynamic	✗ Differs	7.89
Q4 MatMul	✗ Differs	6.83

For better INT8 accuracy, static quantization with Swedish calibration data is recommended.

Export Details

—Optimum version: 2.1.0
—ONNX opset: 17 (native GroupNorm op)
—ORT version: 1.24.4
—Graph: 1838 nodes, single model (no encoder/decoder split)
—Input: input_values — float32, shape [batch, sequence], 16kHz raw waveform
—Output: logits — float32, shape [batch, sequence/320, vocab_size]
—Vocab size: 46 (Swedish character set + CTC blank)
—CTC decoding: not included in graph — greedy argmax or beam search applied in Python
—Note: FP16 model uses keep_io_types=True; I/O remains float32, internal weights are float16

Original Model

See KBLab/wav2vec2-large-voxrex-swedish for training details, dataset, and full evaluation results.

Citation

If you use this model, please cite the original:

bibtex

@misc{wav2vec2-large-voxrex-swedish,
  author = {KBLab},
  title = {wav2vec2-large-voxrex-swedish},
  year = {2022},
  publisher = {HuggingFace},
  url = {https://huggingface.co/KBLab/wav2vec2-large-voxrex-swedish}
}

License

Apache 2.0, same as the original model.