VDNT11 commited on
Commit
c3e8af7
·
verified ·
1 Parent(s): 4283a1b

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +12 -4
app.py CHANGED
@@ -10,18 +10,25 @@ import uuid
10
  device = "cuda:0" if torch.cuda.is_available() else "cpu"
11
  torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
12
 
13
- # Load Whisper model from Hugging Face
14
  @st.cache_resource
15
  def load_model():
16
- model_id = "openai/whisper-large-v2"
17
-
 
 
18
  model = AutoModelForSpeechSeq2Seq.from_pretrained(
19
- model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
 
 
 
20
  )
21
  model.to(device)
22
 
 
23
  processor = AutoProcessor.from_pretrained(model_id)
24
 
 
25
  pipe = pipeline(
26
  "automatic-speech-recognition",
27
  model=model,
@@ -29,6 +36,7 @@ def load_model():
29
  feature_extractor=processor.feature_extractor,
30
  torch_dtype=torch_dtype,
31
  device=device,
 
32
  )
33
  return pipe, processor
34
 
 
10
  device = "cuda:0" if torch.cuda.is_available() else "cpu"
11
  torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
12
 
13
+
14
  @st.cache_resource
15
  def load_model():
16
+ # Use a specific Hindi-optimized Whisper model
17
+ model_id = "openai/whisper-large-v2" # or consider a multilingual model
18
+
19
+ # For Hindi, you might want to specify additional parameters
20
  model = AutoModelForSpeechSeq2Seq.from_pretrained(
21
+ model_id,
22
+ torch_dtype=torch_dtype,
23
+ low_cpu_mem_usage=True,
24
+ use_safetensors=True,
25
  )
26
  model.to(device)
27
 
28
+ # Use the processor from the same model
29
  processor = AutoProcessor.from_pretrained(model_id)
30
 
31
+ # Create pipeline with language specification
32
  pipe = pipeline(
33
  "automatic-speech-recognition",
34
  model=model,
 
36
  feature_extractor=processor.feature_extractor,
37
  torch_dtype=torch_dtype,
38
  device=device,
39
+ generate_kwargs={"language": "hi"} # Specify Hindi language
40
  )
41
  return pipe, processor
42