import torch from transformers import AutoModel, AutoTokenizer repo_id = "humanify/ARAG_embedding_pretrain" device = "cuda" if torch.cuda.is_available() else "cpu" tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True) model = AutoModel.from_pretrained( repo_id, trust_remote_code=True, torch_dtype=torch.float32, ).to(device).eval() model.set_tokenizer(tokenizer) QUERY_INSTRUCTION = "Based on the question asked in the text query and context in the audio query, retrieve the relevant text document associated with that question." DOC_INSTRUCTION = "Represent the user's input." # case 1 # expect [0-1match, 2-3match] # 模型正确区分出性别 # tensor([[0.5530, 0.4304], # [0.6178, 0.4669], # [0.3870, 0.6268], # [0.4139, 0.6821]], device='cuda:0') # tensor([[1.0000, 0.8968, 0.7887, 0.7221], # [0.8968, 1.0000, 0.7567, 0.7600], # [0.7887, 0.7567, 1.0000, 0.9216], # [0.7221, 0.7600, 0.9216, 1.0000]], device='cuda:0') # query_text = [ # "What is the gender of speaker in this audio?", # "What is the gender of speaker in this audio?", # "What is the gender of speaker in this audio?", # "What is the gender of speaker in this audio?", # ] # query_audio = ["./samples/en_male_music.wav", "./samples/en_male.wav", # "./samples/en_female_music.wav", "./samples/en_female.wav"] # case 2 # expect [0-2match, 1-3match] # 测试显示模型是可以区分出是否有背景音乐的 # tensor([[0.4247, 0.1398], # [0.2891, 0.1501], # [0.2446, 0.3100], # [0.2502, 0.3032]], device='cuda:0') # tensor([[1.0000, 0.7550, 0.6976, 0.6746], # [0.7550, 1.0000, 0.7014, 0.6887], # [0.6976, 0.7014, 1.0000, 0.8886], # [0.6746, 0.6887, 0.8886, 1.0000]], device='cuda:0') # query_text = [ # "Describe the background noise of this audio?", # "Describe the background noise of this audio?", # "Describe the background noise of this audio?", # "Describe the background noise of this audio?", # ] # query_audio = ["./samples/en_male_music.wav", "./samples/en_female_music.wav", # "./samples/en_male.wav", "./samples/en_female.wav"] # case 3 # expect [0-2match, 1-3match] # 模型具有quality分辨能力 # tensor([[0.4325, 0.3912], # [0.4801, 0.4459], # [0.4836, 0.4986], # [0.4723, 0.4741]], device='cuda:0') # tensor([[1.0000, 0.9511, 0.9240, 0.8633], # [0.9511, 1.0000, 0.9416, 0.8835], # [0.9240, 0.9416, 1.0000, 0.9293], # [0.8633, 0.8835, 0.9293, 1.0000]], device='cuda:0') # query_text = [ # "Describe the quality of the audio.", # "Describe the quality of the audio.", # "Describe the quality of the audio.", # "Describe the quality of the audio.", # ] # query_audio = ["./samples/en_male_music.wav", "./samples/en_male.wav", # "./samples/en_female_music.wav", "./samples/en_female.wav"] # case 4-1 # expect [1-3match] # query_text = [ # "Describe the speech content and topic.", # "Describe the speech content and topic.", # "Describe the speech content and topic.", # "Describe the speech content and topic.", # ] # query_audio = ["./samples/en_male_sunny.wav", "./samples/en_male.wav", # "./samples/en_female_sunny.wav", "./samples/en_female.wav"] # case 4-2 # expect [1-3match] # 模型可以正确区分出来音频中是否包含某个单词 # tensor([[0.7353, 0.4429], # [0.7130, 0.4250], # [0.5905, 0.6196], # [0.5303, 0.7002]], device='cuda:0') # tensor([[1.0000, 0.9781, 0.8920, 0.7941], # [0.9781, 1.0000, 0.9070, 0.8146], # [0.8920, 0.9070, 1.0000, 0.9668], # [0.7941, 0.8146, 0.9668, 1.0000]], device='cuda:0') # query_text = [ # "Does the speaker mention the word 'sunny' in the audio?", # "Does the speaker mention the word 'sunny' in the audio?", # "Does the speaker mention the word 'sunny' in the audio?", # "Does the speaker mention the word 'sunny' in the audio?", # ] # query_audio = ["./samples/en_male_sunny.wav", "./samples/en_female_sunny.wav", # "./samples/en_male.wav", "./samples/en_female.wav"] # case 5 # expect [0-1 match] # 模型具备区分动物叫声的能力，结果显示如下： # tensor([[0.7022, 0.4406, 0.4651], # [0.7497, 0.3858, 0.4209], # [0.3636, 0.4995, 0.3304], # [0.4793, 0.2937, 0.7720]], device='cuda:0') # tensor([[1.0000, 0.9535, 0.6508, 0.7476], # [0.9535, 1.0000, 0.5688, 0.6673], # [0.6508, 0.5688, 1.0000, 0.5568], # [0.7476, 0.6673, 0.5568, 1.0000]], device='cuda:0') # query_text = [ # "What type of animal is this?", # "What type of animal is this?", # "What type of animal is this?", # "What type of animal is this?", # ] # query_audio = ["./samples/animal_cow_1moo.wav", "./samples/animal_cow_5moo.wav", # "./samples/animal_lion.wav", "./samples/animal_rooster.wav"] # case 6 # expect [0-2 match, 1-3 match] # 模型能区分出中文和英文 # 汉语和英语 # 第一个矩阵结果正确，第二个矩阵有点问题 # tensor([[0.4100, 0.2334], # [0.4451, 0.2417], # [0.3379, 0.3609], # [0.3997, 0.4074]], device='cuda:0') # tensor([[1.0000, 0.9034, 0.8573, 0.7466], # [0.9034, 1.0000, 0.7729, 0.7808], # [0.8573, 0.7729, 1.0000, 0.8199], # [0.7466, 0.7808, 0.8199, 1.0000]], device='cuda:0') # 法语和英语，完全没有问题 # tensor([[0.4100, 0.3098], # [0.4411, 0.3229], # [0.3494, 0.4212], # [0.3523, 0.4130]], device='cuda:0') # tensor([[1.0000, 0.9030, 0.8696, 0.8591], # [0.9030, 1.0000, 0.7874, 0.8468], # [0.8696, 0.7874, 1.0000, 0.9654], # [0.8591, 0.8468, 0.9654, 1.0000]], device='cuda:0') # 英语和西班牙语，完全没有问题 # tensor([[0.4100, 0.3012], # [0.4411, 0.3069], # [0.3274, 0.3698], # [0.3249, 0.3558]], device='cuda:0') # tensor([[1.0000, 0.9030, 0.8591, 0.8590], # [0.9030, 1.0000, 0.7775, 0.8208], # [0.8591, 0.7775, 1.0000, 0.9739], # [0.8590, 0.8208, 0.9739, 1.0000]], device='cuda:0') # 法语和西班牙语，完全没有问题 # tensor([[0.4212, 0.2785], # [0.4170, 0.2737], # [0.2793, 0.3698], # [0.2721, 0.3558]], device='cuda:0') # tensor([[1.0000, 0.9661, 0.8730, 0.8688], # [0.9661, 1.0000, 0.8493, 0.8726], # [0.8730, 0.8493, 1.0000, 0.9739], # [0.8688, 0.8726, 0.9739, 1.0000]], device='cuda:0') # 法语和汉语，完全没问题 # tensor([[0.4212, 0.2278], # [0.4130, 0.2353], # [0.2986, 0.3609], # [0.3602, 0.4074]], device='cuda:0') # tensor([[1.0000, 0.9654, 0.7965, 0.6456], # [0.9654, 1.0000, 0.7990, 0.7011], # [0.7965, 0.7990, 1.0000, 0.8199], # [0.6456, 0.7011, 0.8199, 1.0000]], device='cuda:0') # 西班牙语和汉语，完全没问题 # tensor([[0.3698, 0.1813], # [0.3558, 0.1784], # [0.2814, 0.3609], # [0.3257, 0.4074]], device='cuda:0') # tensor([[1.0000, 0.9739, 0.7871, 0.6249], # [0.9739, 1.0000, 0.7837, 0.6546], # [0.7871, 0.7837, 1.0000, 0.8199], # [0.6249, 0.6546, 0.8199, 1.0000]], device='cuda:0') query_text = [ "Describe the language of the speaker.", "Describe the language of the speaker.", "Describe the language of the speaker.", "Describe the language of the speaker." ] query_audio = ["./samples/spanish_male.wav", "./samples/spanish_female.wav", "./samples/chinese_male.wav", "./samples/zh_female.wav"] # case 7 # expect [0-1 match] + [2-3 match] # 测试结果显示模型不具备区分出说话者语速快慢的能力 # query_text = [ # "Describe the speaking rate of the speaker.", # "Describe the speaking rate of the speaker.", # "Describe the speaking rate of the speaker.", # "Describe the speaking rate of the speaker." # ] # query_audio = ["./samples/female_fast_4.wav", "./samples/female_fast.wav", # "./samples/male_slow.wav", "./samples/female_slow.wav"] # case 8 # expect [0->one, 1->two] # 测试结果显示模型无法区分出声音不同类别的能力 # query_text = [ # "How many distinct sound types are present in this audio?", # "How many distinct sound types are present in this audio?", # "How many distinct sound types are present in this audio?", # "How many distinct sound types are present in this audio?" # ] # query_audio = [ # "./samples/one_type_sound.flac", "./samples/one_type_sound_2.flac", # "./samples/two_type_sound.flac", "./samples/two_type_sound_2.flac", # ] # case 9 # expect [0-1 match] + [2-3 match] # 测试结果显示模型具备对话语类型/语义的判断，结果如下所示。 # tensor([[0.3587, 0.2188], # [0.3793, 0.2482], # [0.2806, 0.3153], # [0.2676, 0.3588]], device='cuda:0') # tensor([[1.0000, 0.9275, 0.6241, 0.5684], # [0.9275, 1.0000, 0.7433, 0.6704], # [0.6241, 0.7433, 1.0000, 0.7269], # [0.5684, 0.6704, 0.7269, 1.0000]], device='cuda:0') # query_text = [ # "Identify the utterance type of this speech.", # "Identify the utterance type of this speech.", # "Identify the utterance type of this speech.", # "Identify the utterance type of this speech." # ] # query_audio = ["./samples/male_statement.wav", "./samples/female_statement.wav", # "./samples/male_question.wav", "./samples/female_question.wav"] # # case 10 # # expect # # 测试结果显示模型可以区分出音频是one-speaker还是two-speaker，结果如下所示。 # # tensor([[0.7003, 0.6864], # # [0.7050, 0.6986], # # [0.3909, 0.7932], # # [0.4326, 0.8110]], device='cuda:0') # # tensor([[1.0000, 0.9369, 0.7105, 0.7634], # # [0.9369, 1.0000, 0.7473, 0.7579], # # [0.7105, 0.7473, 1.0000, 0.8984], # # [0.7634, 0.7579, 0.8984, 1.0000]], device='cuda:0') # query_text = [ # "How many speakers in this audio?", # "How many speakers in this audio?", # "How many speakers in this audio?", # "How many speakers in this audio?", # ] # query_audio = [ # "./samples/en_male.wav", "./samples/en_female.wav", # "./samples/two_speaker.flac", "./samples/two_speaker_2.flac", # ] # case 11-1(说话) # expect # 测试结果如下所示。低音全都是用的male，可能有点不够solid。query-doc矩阵没问题，但是query-query矩阵有问题。 # tensor([[0.4330, 0.4321], # [0.4118, 0.3743], # [0.3756, 0.3932], # [0.4046, 0.4489]], device='cuda:0') # tensor([[1.0000, 0.9168, 0.9511, 0.9505], # [0.9168, 1.0000, 0.9212, 0.9323], # [0.9511, 0.9212, 1.0000, 0.9565], # [0.9505, 0.9323, 0.9565, 1.0000]], device='cuda:0') # query_text = [ # "Is the speaker's tone high or low?", # "Is the speaker's tone high or low?", # "Is the speaker's tone high or low?", # "Is the speaker's tone high or low?" # ] # query_audio = [ # "./samples/female_speech_high_1.wav", "./samples/male_speech_high_1.wav", # "./samples/male_speech_low_1.wav", "./samples/male_speech_low_2.wav" # ] # case 11-2(唱歌) # expect # 测试结果如下所示。低音全都是用的male，可能有点不够solid。 # tensor([[0.4901, 0.3206], # [0.4875, 0.3638], # [0.4354, 0.4760], # [0.4144, 0.4296]], device='cuda:0') # tensor([[1.0000, 0.9399, 0.8666, 0.8854], # [0.9399, 1.0000, 0.9261, 0.9368], # [0.8666, 0.9261, 1.0000, 0.9752], # [0.8854, 0.9368, 0.9752, 1.0000]], device='cuda:0') # query_text = [ # "Is the speaker's tone high or low?", # "Is the speaker's tone high or low?", # "Is the speaker's tone high or low?", # "Is the speaker's tone high or low?" # ] # query_audio = [ # "./samples/female_sing_high_1.wav", "./samples/male_sing_high_1.wav", # "./samples/male_sing_low_1.wav", "./samples/male_sing_low_2.wav" # ] # case 12 # expect [0-1 -> outdoor, 2-3 -> indoor] # 测试结果显示模型无法区分出模型是在室内还是室外发生 # query_text = [ # "Where is this audio most likely recorded?", # "Where is this audio most likely recorded?", # "Where is this audio most likely recorded?", # "Where is this audio most likely recorded?" # ] # query_audio = ["./samples/male_outdoor.wav", "./samples/female_outdoor.wav", # "./samples/indoor.flac", "./samples/indoor2.flac"] # # case 13 # # expect # 测试结果显示模型能够区分出情感差异，结果如下所示。 # tensor([[0.3226, 0.1865], # [0.4463, 0.2482], # [0.4102, 0.4539], # [0.3474, 0.4925]], device='cuda:0') # tensor([[1.0000, 0.7824, 0.5576, 0.5138], # [0.7824, 1.0000, 0.6224, 0.5177], # [0.5576, 0.6224, 1.0000, 0.8121], # [0.5138, 0.5177, 0.8121, 1.0000]], device='cuda:0') # query_text = [ # "What is the emotion of the speaker?", # "What is the emotion of the speaker?", # "What is the emotion of the speaker?", # "What is the emotion of the speaker?" # ] # query_audio = [ # "./samples/people_happy_1.flac", "./samples/people_happy_2.flac", # "./samples/people_cry_1.flac", "./samples/female_sad_1.wav" # ] # # case 14 # # expect [0->child, 1->adult] # # 测试结果如下所示，显示模型具备能力区分说话者的大致年龄。但是两个young都是用的boy； # # 如果用young_girl.wav的话, 第一个矩阵没问题，第二个矩阵有问题。 # # tensor([[0.5183, 0.5019], # # [0.5069, 0.4590], # # [0.2864, 0.5422], # # [0.3039, 0.5131]], device='cuda:0') # # tensor([[1.0000, 0.9227, 0.7293, 0.7653], # # [0.9227, 1.0000, 0.6910, 0.7291], # # [0.7293, 0.6910, 1.0000, 0.8695], # # [0.7653, 0.7291, 0.8695, 1.0000]], device='cuda:0') # query_text = [ # "Estimate the age of the speaker.", # "Estimate the age of the speaker.", # "Estimate the age of the speaker.", # "Estimate the age of the speaker." # ] # query_audio = [ # "./samples/young_boy_1.flac", "./samples/young_boy_2.wav", # "./samples/adult_man.flac", "./samples/adult_woman.flac" # ] # case 15 # 测试结果显示模型可以区分出不同的英语口音，结果如下所示。其中第二个矩阵中有点不够solid # tensor([[0.7703, 0.7582], # [0.7267, 0.7185], # [0.7237, 0.7307], # [0.7646, 0.7759]], device='cuda:0') # tensor([[1.0000, 0.9324, 0.9442, 0.9598], # [0.9324, 1.0000, 0.9827, 0.9704], # [0.9442, 0.9827, 1.0000, 0.9857], # [0.9598, 0.9704, 0.9857, 1.0000]], device='cuda:0') # query_text = [ # "Identify the English accent of the speaker.", # "Identify the English accent of the speaker.", # "Identify the English accent of the speaker.", # "Identify the English accent of the speaker." # ] # query_audio = ["./samples/male_america_english.wav", "./samples/female_america_english.wav", # "./samples/male_british_english.wav", "./samples/female_british_english.wav"] # # case 16 # # expect # # 测试结果显示模型可以区分出飞机引擎声和大巴引擎声，结果如下所示 # # tensor([[0.4200, 0.2822], # # [0.5744, 0.2713], # # [0.2183, 0.3696], # # [0.2527, 0.3816]], device='cuda:0') # # tensor([[1.0000, 0.8755, 0.7497, 0.7796# [0.6261, 0.5706, 1.0000, 0.7327], # [0.4892, 0.5031, 0.7327, 1.0000]], device='cuda:0') # query_text = [ # "What is the source of this sound?", # "What is the source of this sound?", # "What is the source of this sound?", # "What is the source of this sound?" # ] # query_audio = ["./samples/rainy.wav", "./samples/rainy_2.flac", # "./samples/wind_2.flac", "./samples/wind.flac"] # case 18 # expect # 测试结果显示模型具备对说话者的计数能力； # 3-speaker/4-speaker测试结果如下, query-doc矩阵没问题，query-query矩阵有点小问题 # tensor([[0.4325, 0.3912], # [0.4801, 0.4459], # [0.4836, 0.4986], # [0.4723, 0.4741]], device='cuda:0') # tensor([[1.0000, 0.9511, 0.9240, 0.8633], # [0.9511, 1.0000, 0.9416, 0.8835], # [0.9240, 0.9416, 1.0000, 0.9293], # [0.8633, 0.8835, 0.9293, 1.0000]], device='cuda:0') # query_text = [ # "What's the number of participants in the current conversation?", # "What's the number of participants in the current conversation?", # "What's the number of participants in the current conversation?", # "What's the number of participants in the current conversation?" # ] # query_audio = [ # "./samples/three_speaker_2.wav", "./samples/three_speaker.wav", # "./samples/four_speaker.wav", "./samples/four_speaker_3.wav", # ] # case 19 # 测试结果显示不具备音乐风格识别能力 # query_text = [ # "What is the music style in the audio?", # "What is the music style in the audio?", # "What is the music style in the audio?", # "What is the music style in the audio?" # ] # query_audio = [ # "./samples/hiphop_1.flac", "./samples/hiphop_2.flac", # "./samples/classical_1.flac", "./samples/classical_2.flac", # ] # case 20 # 测试结果显示模型具备识别乐器声音的能力 # tensor([[0.4882, 0.3500], # [0.4124, 0.3477], # [0.3047, 0.6171], # [0.4425, 0.5910]], device='cuda:0') # tensor([[1.0000, 0.8528, 0.7110, 0.8001], # [0.8528, 1.0000, 0.7338, 0.7728], # [0.7110, 0.7338, 1.0000, 0.9112], # [0.8001, 0.7728, 0.9112, 1.0000]], device='cuda:0') # query_text = [ # "What musical instrument is producing the sound?", # "What musical instrument is producing the sound?", # "What musical instrument is producing the sound?", # "What musical instrument is producing the sound?" # ] # query_audio = [ # "./samples/piano_3.flac", "./samples/piano_2.flac", # "./samples/guitar_1.flac", "./samples/guitar_2.flac", # ] # case 21 # 测试显示模型无法计数说话中单词的数量 # query_text = [ # "How many times does the word 'said' appear in the audio?", # "How many times does the word 'beautiful' appear in the audio?", # "How many times does the word 'fate' appear in the audio?", # "How many times does the word 'ninty' appear in the audio?" # ] # query_audio = [ # "./samples/1_said.wav", "./samples/1_beautiful.wav", # "./samples/1_fate.wav", "./samples/3_ninty_nine.wav", # ] # case 22 # 测试显示模型是否具备计数铃声和鼓声的能力 # query_text = [ # "How many times does the sound appear in the audio?", # "How many times does the sound appear in the audio?", # "How many times does the sound appear in the audio?", # "How many times does the sound appear in the audio?" # ] # query_audio = [ # "./samples/2_drumbeat.wav", "./samples/2_ringtone.wav", # "./samples/6_drumbeat.wav", "./samples/6_ringtone.wav" # ] # ---- # doc_text = [ # "male", # "female", # ] # doc_text = [ # "music", # "nothing", # ] # doc_text = [ # "High-quality", # "Low-quality", # ] # doc_text = [ # "yes", # "no", # ] # doc_text = [ # "cow", # "lion", # "rooster", # ] doc_text = [ "spanish", "chinese", ] # doc_text = [ # "fast", # "slow", # ] # doc_text = [ # "one type", # "two types", # ] # doc_text = [ # "statement", # "question", # ] # doc_text = [ # "one speaker", # "two speakers", # ] # doc_text = [ # "high", # "low", # ] # doc_text = [ # "outdoor", # "indoor", # ] # doc_text = [ # "happy", # "sad", # ] # doc_text = [ # "child", # "adult", # ] # doc_text = [ # "American accent", # "British accent", # ] # doc_text = [ # "helicopter engine", # "car engine", # ] # doc_text = [ # "rain", # "wind", # ] # doc_text = [ # "hiphop", # "classical", # ] # doc_text = [ # "three persons", # "four persons", # ] # doc_text = [ # "piano", # "guitar", # ] # doc_text = [ # "one", # "three", # ] # doc_text = [ # "two", # "six", # ] query_embeddings = model.encode( text=query_text, audio=query_audio, task="query", instruction=QUERY_INSTRUCTION, normalize=True, device=device, ) doc_embeddings = model.encode( text=doc_text, # text=[None, None], # audio=doc_audio, task="document", instruction=DOC_INSTRUCTION, normalize=True, device=device, ) similarity = query_embeddings @ doc_embeddings.T print(similarity) similarity = query_embeddings @ query_embeddings.T print(similarity)