File size: 4,765 Bytes
4f3f83c
425758f
4f3f83c
 
ec1a913
 
 
 
 
6769f1b
ec1a913
4f3f83c
425758f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4f3f83c
8299dc1
 
4f3f83c
425758f
4f3f83c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6769f1b
 
 
 
 
 
 
 
 
 
4f3f83c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aabfdd7
4f3f83c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4831744
4f3f83c
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from transformers.pipelines.audio_utils import ffmpeg_read
import gradio as gr
from transformers import WhisperForConditionalGeneration, WhisperProcessor
from transformers.models.whisper.tokenization_whisper import LANGUAGES
from transformers.pipelines.audio_utils import ffmpeg_read

model_id = "openai/whisper-large-v2"
device = "cuda" if torch.cuda.is_available() else "CPU"


LANGUANGE_MAP = {
    0: 'Arabic',
    1: 'Basque',
    2: 'Breton',
    3: 'Catalan',
    4: 'Chinese_China',
    5: 'Chinese_Hongkong',
    6: 'Chinese_Taiwan',
    7: 'Chuvash',
    8: 'Czech',
    9: 'Dhivehi',
    10: 'Dutch',
    11: 'English',
    12: 'Esperanto',
    13: 'Estonian',
    14: 'French',
    15: 'Frisian',
    16: 'Georgian',
    17: 'German',
    18: 'Greek',
    19: 'Hakha_Chin',
    20: 'Indonesian',
    21: 'Interlingua',
    22: 'Italian',
    23: 'Japanese',
    24: 'Kabyle',
    25: 'Kinyarwanda',
    26: 'Kyrgyz',
    27: 'Latvian',
    28: 'Maltese',
    29: 'Mongolian',
    30: 'Persian',
    31: 'Polish',
    32: 'Portuguese',
    33: 'Romanian',
    34: 'Romansh_Sursilvan',
    35: 'Russian',
    36: 'Sakha',
    37: 'Slovenian',
    38: 'Spanish',
    39: 'Swedish',
    40: 'Tamil',
    41: 'Tatar',
    42: 'Turkish',
    43: 'Ukranian',
    44: 'Welsh'
 }





model_ckpt = "barto17/language-detection-fine-tuned-on-xlm-roberta-base"
model = AutoModelForSequenceClassification.from_pretrained(model_ckpt)
tokenizer = AutoTokenizer.from_pretrained(model_ckpt)

def detect_language(sentence):
    tokenized_sentence = tokenizer(sentence, return_tensors='pt')
    output = model(**tokenized_sentence)
    predictions = torch.nn.functional.softmax(output.logits, dim=-1)
    probability, pred_idx = torch.max(predictions, dim=-1)
    language = LANGUANGE_MAP[pred_idx.item()]
    return language, probability.item()


def process_audio_file(file):
    with open(file, "rb") as f:
        inputs = f.read()

    audio = ffmpeg_read(inputs, sampling_rate)
    return audio

def transcribe(Microphone, File_Upload):
    processor = WhisperProcessor.from_pretrained(model_id)
    model = WhisperForConditionalGeneration.from_pretrained(model_id)
    model.eval()
    model.to(device)

    sampling_rate = processor.feature_extractor.sampling_rate
    
    bos_token_id = processor.tokenizer.all_special_ids[-106]
    decoder_input_ids = torch.tensor([bos_token_id]).to(device)

    warn_output = ""
    if (Microphone is not None) and (File_Upload is not None):
        warn_output = "WARNING: You've uploaded an audio file and used the microphone. " \
                      "The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
        file = Microphone

    elif (Microphone is None) and (File_Upload is None):
        return "ERROR: You have to either use the microphone or upload an audio file"

    elif Microphone is not None:
        file = Microphone
    else:
        file = File_Upload

    audio_data = process_audio_file(file)

    input_features = processor(audio_data, return_tensors="pt").input_features
    
    with torch.no_grad():
        logits = model.forward(input_features.to(device), decoder_input_ids=decoder_input_ids).logits
    
    pred_ids = torch.argmax(logits, dim=-1)
    transcription = processor.decode(pred_ids[0])
    
    detect_language(transcription.capitalize())


examples=['sample1.mp3', 'sample2.mp3', 'sample3.mp3']
examples = [[f"./{f}"] for f in examples]

outputs=gr.outputs.Label(label="Language detected:")
article = """
Fine-tuned on xlm-roberta-base model.\n
Supported languages:\n 
    'Arabic', 'Basque', 'Breton', 'Catalan', 'Chinese_China', 'Chinese_Hongkong', 'Chinese_Taiwan', 'Chuvash', 'Czech', 
    'Dhivehi', 'Dutch', 'English', 'Esperanto', 'Estonian', 'French', 'Frisian', 'Georgian', 'German', 'Greek', 'Hakha_Chin', 
    'Indonesian', 'Interlingua', 'Italian', 'Japanese', 'Kabyle', 'Kinyarwanda', 'Kyrgyz', 'Latvian', 'Maltese', 
    'Mangolian', 'Persian', 'Polish', 'Portuguese', 'Romanian', 'Romansh_Sursilvan', 'Russian', 'Sakha', 'Slovenian', 
    'Spanish', 'Swedish', 'Tamil', 'Tatar', 'Turkish', 'Ukranian', 'Welsh'
"""

gr.Interface(
    fn=transcribe,
    inputs=[
        gr.inputs.Audio(source="microphone", type='filepath', optional=True),
        gr.inputs.Audio(source="upload", type='filepath', optional=True),
    ],
    
    outputs=[
        gr.outputs.Textbox(label="Language"),
        gr.Number(label="Probability"),
    ],

    verbose=True,
    examples = examples,
    title="Language Identification from Audio",
    description="Detect the Language from Audio.",
    article=article,
    theme="huggingface"
).launch()