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 noise we actually provide a split labeled with the original or "natural" correct transcriptions, and one labeled with our selected "target" transcriptions. For instance, the split contains the modified original transcription, while the An ASR model that this dataset fools would get a low target WER and a high natural WER. An ASR model robust to this dataset would get a low natural WER and a high target WER. Therefore we provide in total 8 splits:

natural_nat_txt
natural_tgt_txt
adv_0.04_nat_txt
adv_0.04_tgt_txt
adv_0.015_nat_txt
adv_0.015_tgt_txt
adv_0.015_RIR_nat_txt
adv_0.015_RIR_tgt_txt

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:", wer(result["text"], result["transcription"]))

Result (WER):

"0.015 target text"	"0.015 natural text"	"0.04 target text"	"0.04 natural text"
58.2	108	49.5	108