Lauler's picture
citation info
1b982bc
|
raw
history blame
3.46 kB
metadata
language: sv
arxiv: https://arxiv.org/abs/2205.03026
datasets:
  - common_voice
  - NST_Swedish_ASR_Database
  - P4
metrics:
  - wer
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - hf-asr-leaderboard
license: cc0-1.0
model-index:
  - name: Wav2vec 2.0 large VoxRex Swedish
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice
          type: common_voice
          args: sv-SE
        metrics:
          - name: Test WER
            type: wer
            value: 8.49

Wav2vec 2.0 large VoxRex Swedish (C)

Disclaimer: This is a work in progress. See VoxRex for more details.

Update 2022-01-10: Updated to VoxRex-C version.

Update 2022-05-16: Paper is is here.

Finetuned version of KBs VoxRex large model using Swedish radio broadcasts, NST and Common Voice data. Evalutation without a language model gives the following: WER for NST + Common Voice test set (2% of total sentences) is 2.5%. WER for Common Voice test set is 8.49% directly and 7.37% with a 4-gram language model.

When using this model, make sure that your speech input is sampled at 16kHz.

Performance*

Comparison

*Chart shows performance without the additional 20k steps of Common Voice fine-tuning

Training

This model has been fine-tuned for 120000 updates on NST + CommonVoice and then for an additional 20000 updates on CommonVoice only. The additional fine-tuning on CommonVoice hurts performance on the NST+CommonVoice test set somewhat and, unsurprisingly, improves it on the CommonVoice test set. It seems to perform generally better though [citation needed].

WER during training

Usage

The model can be used directly (without a language model) as follows:

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
test_dataset = load_dataset("common_voice", "sv-SE", split="test[:2%]").
processor = Wav2Vec2Processor.from_pretrained("KBLab/wav2vec2-large-voxrex-swedish")
model = Wav2Vec2ForCTC.from_pretrained("KBLab/wav2vec2-large-voxrex-swedish")
resampler = torchaudio.transforms.Resample(48_000, 16_000)
# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch
test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
with torch.no_grad():
    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])

Citation

https://arxiv.org/abs/2205.03026

@misc{malmsten2022hearing,
      title={Hearing voices at the National Library -- a speech corpus and acoustic model for the Swedish language}, 
      author={Martin Malmsten and Chris Haffenden and Love Börjeson},
      year={2022},
      eprint={2205.03026},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}