(You must get at least 85% to be considered the same person)
"""
)
EFFECTS = [
["remix", "-"],
["channels", "1"],
["rate", "16000"],
["gain", "-1.0"],
["silence", "1", "0.1", "0.1%", "-1", "0.1", "0.1%"],
["trim", "0", "10"],
]
THRESHOLD = 0.85
model_name = "microsoft/unispeech-sat-base-plus-sv"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
model = AutoModelForAudioXVector.from_pretrained(model_name).to(device)
cosine_sim = torch.nn.CosineSimilarity(dim=-1)
def similarity_fn(path1, path2):
if not (path1 and path2):
return '
ERROR: Please record audio for *both* speakers!'
wav1, _ = apply_effects_file(path1, EFFECTS)
wav2, _ = apply_effects_file(path2, EFFECTS)
print(wav1.shape, wav2.shape)
input1 = feature_extractor(wav1.squeeze(0), return_tensors="pt", sampling_rate=16000).input_values.to(device)
input2 = feature_extractor(wav2.squeeze(0), return_tensors="pt", sampling_rate=16000).input_values.to(device)
with torch.no_grad():
emb1 = model(input1).embeddings
emb2 = model(input2).embeddings
emb1 = torch.nn.functional.normalize(emb1, dim=-1).cpu()
emb2 = torch.nn.functional.normalize(emb2, dim=-1).cpu()
similarity = cosine_sim(emb1, emb2).numpy()[0]
if similarity >= THRESHOLD:
output = OUTPUT_OK.format(similarity * 100)
else:
output = OUTPUT_FAIL.format(similarity * 100)
return output
inputs = [
gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker #1"),
gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker #2"),
]
output = gr.outputs.HTML(label="")
description = (
"This demo from Microsoft will compare two speech samples and determine if they are from the same speaker. "
"Try it with your own voice! If you find an incorrect prediction, you can click FLAG to save the recordings to a public dataset: "
"https://huggingface.co/datasets/abidlabs/voice-verification-adversarial-dataset, "
"consisting of samples on which the model makes mistakes, which may further improve research in this field. Disclaimer: this will "
"save the recordings to a PUBLIC dataset so please be careful about what you FLAG."
)
article = (
"
"
"🎙️ Learn more about UniSpeech-SAT | "
"📚 UniSpeech-SAT paper | "
"📚 X-Vector paper"
"
"
)
examples = [
["samples/cate_blanch.mp3", "samples/cate_blanch_2.mp3"],
["samples/cate_blanch.mp3", "samples/kirsten_dunst.wav"],
]
HF_TOKEN = os.getenv('HF_TOKEN')
hf_saver = gr.HuggingFaceDatasetSaver(HF_TOKEN, "voice-verification-adversarial-dataset")
interface = gr.Interface(
fn=similarity_fn,
inputs=inputs,
outputs=output,
description=description,
title="Break this voice verification model!",
layout="horizontal",
theme="huggingface",
live=False,
examples=examples,
article="[Link to dataset](https://huggingface.co/datasets/abidlabs/voice-verification-adversarial-dataset)",
)
interface.launch(enable_queue=True)