or4cl3ai/SoundSlayerAI
Text-to-Speech
•
Updated
•
28
•
16
The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider
removing the
loading script
and relying on
automated data support
(you can use
convert_to_parquet
from the datasets
library). If this is not possible, please
open a discussion
for direct help.
This dataset is a subset of LibriSpeech and Multilingual CommonVoice that have been adversarially modified to fool Whisper ASR model.
Original source code.
The raw tar files.
targeted
config contains targeted adversarial examples. When successful, they fool Whisper into predicting the sentence OK Google, browse to evil.com
even if the input is entirely different. We provide a split for each Whisper model, and one containing the original, unmodified inputsuntargeted-35
and untargeted-40
configs contain untargeted adversarial examples, with average Signal-Noise Ratios of 35dB and 40dB respectively. They fool Whisper into predicting erroneous transcriptions. We provide a split for each Whisper model, and one containing the original, unmodified inputslanguage-<lang> configs contain adversarial examples in language <lang> that fool Whisper in predicting the wrong language. Split
.contain inputs that Whisper perceives as <target_lang>, and split
.original` contains the original inputs in language . We use 3 target languages (English, Tagalog and Serbian) and 7 source languages (English, Italian, Indonesian, Danish, Czech, Lithuanian and Armenian).Here is an example of code using this dataset:
model_name="whisper-medium"
config_name="targeted"
split_name="whisper.medium"
hub_path = "openai/whisper-"+model_name
processor = WhisperProcessor.from_pretrained(hub_path)
model = WhisperForConditionalGeneration.from_pretrained(hub_path).to("cuda")
dataset = load_dataset("RaphaelOlivier/whisper_adversarial_examples",config_name ,split=split_name)
def map_to_pred(batch):
input_features = processor(batch["audio"][0]["array"], return_tensors="pt").input_features
predicted_ids = model.generate(input_features.to("cuda"))
transcription = processor.batch_decode(predicted_ids, normalize = True)
batch['text'][0] = processor.tokenizer._normalize(batch['text'][0])
batch["transcription"] = transcription
return batch
result = dataset.map(map_to_pred, batched=True, batch_size=1)
wer = load("wer")
for t in zip(result["text"],result["transcription"]):
print(t)
print(wer.compute(predictions=result["text"], references=result["transcription"]))