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

The speakers are

{:.1f}%

similar

Welcome, human!

(You must get at least 85% to be considered the same person)
""" ) OUTPUT_FAIL = ( """

The speakers are

{:.1f}%

similar

You shall not pass!

(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.Audio(sources=["microphone"], type="filepath", label="Speaker #1"), gr.Audio(sources=["microphone"], type="filepath", label="Speaker #2"), ] output = gr.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!" ) 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/heath_ledger.mp3"], ] interface = gr.Interface( fn=similarity_fn, inputs=inputs, outputs=output, layout="horizontal", allow_flagging="never", live=False, examples=examples, cache_examples=False ) interface.launch()