RaphaelOlivier/librispeech_asr_adversarial

Description

This dataset is a subset of https://huggingface.co/datasets/librispeech_asr that has been adversarially modified. It is designed to fool ASR models into predicting a target of our choosing instead of the correct output.

Splits

The dataset contains several splits. Each split consists of the same utterances, modified with different types and amount of noise. 3 noises have been used:

• Adversarial noise of radius 0.04 (adv_0.04 split)
• Adversarial noise of radius 0.015 (adv_0.015 split)
• Adversarial noise of radius 0.015 combined with Room Impulse Response (RIR) noise (adv_0.015_RIR split)

In addition we provide the original inputs (natural split)

For each split we actually provide two text keys: true_text which is the original LibriSpeech label, i.e. the sentence one can actually hear when listening to the audio; and target_text, which is the target sentence of our adversarial attack. An ASR model that this dataset fools would get a low WER on target_text and a high WER on true_text. An ASR model robust to this dataset would get the opposite.

Usage

You should evaluate your model on this dataset as you would evaluate it on LibriSpeech. Here is an example with Wav2Vec2

from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import torch
from jiwer import wer


librispeech_adv_eval = load_dataset("RaphaelOlivier/librispeech_asr_adversarial", "adv", split="adv_0.15_adv_txt")

model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h").to("cuda")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")

def map_to_pred(batch):
    input_values = processor(batch["audio"]["array"], return_tensors="pt", padding="longest").input_values
    with torch.no_grad():
        logits = model(input_values.to("cuda")).logits

    predicted_ids = torch.argmax(logits, dim=-1)
    transcription = processor.batch_decode(predicted_ids)
    batch["transcription"] = transcription
    return batch

result = librispeech_adv_eval.map(map_to_pred, batched=True, batch_size=1, remove_columns=["audio"])

print("WER on correct labels:", wer(result["true_text"], result["transcription"]))
print("WER on attack targets:", wer(result["target_text"], result["transcription"]))

Result (WER):

"0.015 target_text"	"0.015 true_text"	"0.04 target_text"	"0.04 true_text"
58.2	108	49.5	108

License

We release this dataset under the CC By 4.0 license, similarly to the original LibriSpeech dataset.