Anmolkhurana88/Sign-Language-Translator

1from google import genai
2from collections import deque
3import time
4
5from dotenv import load_dotenv
6import os
7
8load_dotenv()
9
10client = genai.Client()
11
12# generator = pipeline(task='text-generation', model="meta-llama/Llama-3.2-3B-Instruct")
13
14def generator(prompt, max_new_tokens=50):
15    try:
16      response = client.models.generate_content(
17        model="gemini-2.5-flash",
18        contents=prompt
19      )
20      return response.text
21    
22    except Exception as e:
23      print(f"Error in text generation: {e}")
24      return ""
25
26def generate_text(gloss_input, last_text=''):
27    instruct = 'You are a gloss-to-English converter. Output only the sentence using only given gloss tokens. No need to complete it with additional words. No explanations.'
28    prompt = f'{instruct}\nGloss: {gloss_input}\nSentence:'
29
30    max_tokens = len(gloss_input.split()) + len(prompt.split())
31
32    start = time.time()
33    output = generator(prompt, max_new_tokens=max_tokens)
34    # output = gloss_input
35    
36    end = time.time()
37    print(f"Time taken for generation: {end - start} seconds")
38
39    if prompt in output:
40        output = output.replace(prompt, '').strip()
41    if '(' in output:
42        output = output.split('(')[0].strip()
43    else:
44        output = output.strip()
45
46    output = output.replace('_', ' ').strip()
47    return output.split('\n')[0]
48
49WINDOW_SIZE = 8
50MIN_TRIGGER = 4
51MAX_CONSUME = 6
52SILENCE_TIMEOUT = 3
53CONF_THRESHOLD = 0.7
54
55# Buffer Manager Class
56class GlossBuffer:
57  def __init__(self):
58    self.buffer = deque(maxlen=WINDOW_SIZE)
59    self.last_gloss_time = 0
60
61  def append_gloss(self, gloss):
62    curr_time = time.time()
63    self.update_buffer(curr_time)
64
65    if gloss not in self.get_buffer():
66      self.buffer.append((gloss, curr_time))
67
68    self.last_gloss_time = curr_time
69
70  def update_buffer(self, curr_time):
71    curr_time = time.time()
72
73    if curr_time - self.last_gloss_time > SILENCE_TIMEOUT:
74      # Clear buffer after prolonged silence
75      self.buffer.clear()
76
77    # Remove old glosses
78    while len(self.buffer) > 0 and self.buffer[0][1] < curr_time - SILENCE_TIMEOUT:
79      self.buffer.popleft()
80
81  def get_buffer(self):
82    return [t for t,c in self.buffer]
83
84  def get_gloss_list(self, counter):
85    gloss_list = self.get_buffer()
86
87    if len(gloss_list) < MIN_TRIGGER and time.time() - counter['last_text_time'] < SILENCE_TIMEOUT:
88      # Not enough glosses to trigger generation and recently generated text
89      return []
90    
91    counter['last_text_time'] = time.time()
92
93    # removing older glosses
94    if len(gloss_list) > MAX_CONSUME:
95      gloss_list = gloss_list[:MAX_CONSUME]
96
97    return gloss_list
98  
99
100def generate_continue_text(gloss_buffer, text_buffer, counter):
101  gloss_list = gloss_buffer.get_gloss_list(counter)
102
103  if len(gloss_list) > 0:
104    text_list = list(text_buffer)
105
106    gloss_text = ' '.join(gloss_list)
107    
108    if len(gloss_list) >= MIN_TRIGGER:
109      gen_text = generate_text(gloss_text, ' '.join(text_list))
110    else:
111      gen_text = gloss_text
112
113    if gen_text and gen_text.strip() != '':
114      text_buffer.extend(gen_text.split())
115      return gen_text
116    
117  return ''
118
119def create_text_buffer(max_size=50):
120    return deque(maxlen=max_size)
121
122if __name__ == "__main__":
123    text = "HELLO HOW YOU/YOUR FEEL TODAY I/ME THINK FUTURE CAREER PLAN YOU/YOUR LIKE/LOVE BOOK_READ OR MOVIE/FILM"
124    gloss_buffer = GlossBuffer()
125
126    text_buffer = create_text_buffer()
127
128    for word in text.split():
129      gloss_buffer.append_gloss(word)
130      generate_continue_text(gloss_buffer, text_buffer, {})