import torch import gradio as gr from torchaudio.sox_effects import apply_effects_file from transformers import AutoFeatureExtractor, AutoModelForAudioXVector device = torch.device("cuda" if torch.cuda.is_available() else "cpu") STYLE = """ """ OUTPUT_OK = STYLE + """

The speakers are

{:.1f}%

similar

Welcome, human!

(You must get 89% or more to be considered the same person)
""" OUTPUT_FAIL = STYLE + """

The speakers are

{:.1f}%

similar

You shall not pass!

(You must get 89% or more to be considered the same person)
""" EFFECTS = [ ["channels", "1"], ["rate", "16000"], ["gain", "-3.0"], ["silence", "1", "0.1", "0.1%", "-1", "0.1", "0.1%"], ] THRESHOLD = 0.89 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(mic_path1, file_path1, mic_path2, file_path2): if not ((mic_path1 or file_path1) and (mic_path2 or file_path2)): return 'ERROR: Please record or upload audio for *both* speakers!' wav1, _ = apply_effects_file(mic_path1 if mic_path1 else file_path1, EFFECTS) wav2, _ = apply_effects_file(mic_path2 if mic_path2 else 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="upload", type="filepath", optional=True, label="or"), gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker #2"), gr.inputs.Audio(source="upload", type="filepath", optional=True, label="or"), ] output = gr.outputs.HTML(label="") description = ( "Speaker Verification demo based on " "UniSpeech-SAT: Universal Speech Representation Learning with Speaker Aware Pre-Training" ) article = ( "

" "🎙️ Learn more about UniSpeech-SAT | " "📚 UniSpeech-SAT paper | " "📚 X-Vector paper" "

" ) interface = gr.Interface( fn=similarity_fn, inputs=inputs, outputs=output, title="Speaker Verification with UniSpeech-SAT + X-Vectors", description=description, article=article, layout="horizontal", theme="huggingface", allow_flagging=False, live=False, ) interface.launch(enable_queue=True)