Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -3,6 +3,7 @@ from transformers import WhisperProcessor, WhisperForConditionalGeneration
|
|
3 |
import torch
|
4 |
import torchaudio
|
5 |
import spaces
|
|
|
6 |
|
7 |
# Initialize devices
|
8 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
@@ -12,13 +13,78 @@ processor = WhisperProcessor.from_pretrained("aiola/whisper-ner-v1")
|
|
12 |
model = WhisperForConditionalGeneration.from_pretrained("aiola/whisper-ner-v1")
|
13 |
model = model.to(device)
|
14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
def unify_ner_text(text, symbols_to_replace=("/", " ", ":", "_")):
|
16 |
"""Process and standardize entity text by replacing certain symbols and normalizing spaces."""
|
17 |
-
text = " ".join(text.split())
|
18 |
for symbol in symbols_to_replace:
|
19 |
-
text = text.replace(symbol, "-")
|
20 |
return text.lower()
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
@spaces.GPU # This decorator ensures your function can use GPU on Hugging Face Spaces
|
23 |
def transcribe_and_recognize_entities(audio_file, prompt):
|
24 |
target_sample_rate = 16000
|
@@ -48,14 +114,56 @@ def transcribe_and_recognize_entities(audio_file, prompt):
|
|
48 |
)
|
49 |
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
|
50 |
|
51 |
-
|
|
|
|
|
52 |
|
53 |
-
iface = gr.Interface(
|
54 |
-
fn=transcribe_and_recognize_entities,
|
55 |
-
inputs=[gr.Audio(label="Upload Audio", type="filepath"), gr.Textbox(label="Entity Recognition Prompt")],
|
56 |
-
outputs=gr.Textbox(label="Transcription and Entities"),
|
57 |
-
title="Whisper-NER Demo",
|
58 |
-
description="Upload an audio file and enter entities to identify. The model will transcribe the audio and recognize entities."
|
59 |
-
)
|
60 |
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
import torch
|
4 |
import torchaudio
|
5 |
import spaces
|
6 |
+
import re
|
7 |
|
8 |
# Initialize devices
|
9 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
13 |
model = WhisperForConditionalGeneration.from_pretrained("aiola/whisper-ner-v1")
|
14 |
model = model.to(device)
|
15 |
|
16 |
+
|
17 |
+
examples = [
|
18 |
+
[
|
19 |
+
"audio/672-122797-0026.wav",
|
20 |
+
"monetary-value, biological-classification, desire, demographic-group, object-category, relationship-role, reflexive-pronoun, furniture-type"
|
21 |
+
],
|
22 |
+
[
|
23 |
+
"audio/672-122797-0024.wav",
|
24 |
+
"health-warning, importance-indicator, event, sentiment"
|
25 |
+
],
|
26 |
+
[
|
27 |
+
"audio/672-122797-0027.wav",
|
28 |
+
"action, emotional-resilience, comparative-path-characteristic, social-role"
|
29 |
+
],
|
30 |
+
[
|
31 |
+
"audio/672-122797-0048.wav",
|
32 |
+
"weapon, emotional-state, household-chore, atmosphere-quality"
|
33 |
+
],
|
34 |
+
[
|
35 |
+
"audio/7021-85628-0025.wav",
|
36 |
+
"action-goal, person's-title, emotional-connection, personal-qualities, pronoun-target, assignmentaction, physical-action, family-role"
|
37 |
+
]
|
38 |
+
]
|
39 |
+
|
40 |
+
|
41 |
def unify_ner_text(text, symbols_to_replace=("/", " ", ":", "_")):
|
42 |
"""Process and standardize entity text by replacing certain symbols and normalizing spaces."""
|
43 |
+
text = " ".join(text.split())
|
44 |
for symbol in symbols_to_replace:
|
45 |
+
text = text.replace(symbol, "-")
|
46 |
return text.lower()
|
47 |
|
48 |
+
|
49 |
+
def extract_entities_and_clean_text_fixed(text):
|
50 |
+
entity_pattern = r"<(.*?)>(.*?)<\1>>"
|
51 |
+
entities = []
|
52 |
+
clean_text = []
|
53 |
+
current_pos = 0
|
54 |
+
|
55 |
+
# Iterate through the matches for entity tags
|
56 |
+
for match in re.finditer(entity_pattern, text):
|
57 |
+
# Add text before the entity to the clean text
|
58 |
+
clean_text.append(text[current_pos:match.start()])
|
59 |
+
|
60 |
+
entity_type = match.group(1)
|
61 |
+
entity_text = match.group(2)
|
62 |
+
start_pos = len("".join(clean_text)) # Start position in the clean text
|
63 |
+
end_pos = start_pos + len(entity_text)
|
64 |
+
|
65 |
+
# Append the entity text to the clean text
|
66 |
+
clean_text.append(entity_text)
|
67 |
+
|
68 |
+
# Add the entity details to the list
|
69 |
+
entities.append({
|
70 |
+
"entity": entity_type,
|
71 |
+
"text": entity_text,
|
72 |
+
"start": start_pos,
|
73 |
+
"end": end_pos
|
74 |
+
})
|
75 |
+
|
76 |
+
# Update the current position to the end of the match
|
77 |
+
current_pos = match.end()
|
78 |
+
|
79 |
+
# Append the remaining part of the text after the last entity
|
80 |
+
clean_text.append(text[current_pos:])
|
81 |
+
|
82 |
+
# Join all parts of the clean text
|
83 |
+
clean_text_str = "".join(clean_text)
|
84 |
+
|
85 |
+
return clean_text_str, entities
|
86 |
+
|
87 |
+
|
88 |
@spaces.GPU # This decorator ensures your function can use GPU on Hugging Face Spaces
|
89 |
def transcribe_and_recognize_entities(audio_file, prompt):
|
90 |
target_sample_rate = 16000
|
|
|
114 |
)
|
115 |
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
|
116 |
|
117 |
+
clean_text_fixed, extracted_entities_fixed = extract_entities_and_clean_text_fixed(transcription)
|
118 |
+
|
119 |
+
return transcription, {"text": clean_text_fixed, "entities": extracted_entities_fixed}
|
120 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
121 |
|
122 |
+
with gr.Blocks(title="WhisperNER v1") as demo:
|
123 |
+
|
124 |
+
gr.Markdown(
|
125 |
+
"""
|
126 |
+
# Whisper-NER: ASR with zero-shot NER
|
127 |
+
|
128 |
+
WhisperNER is a unified model for automatic speech recognition (ASR) and named entity recognition (NER), with zero-shot capabilities.
|
129 |
+
|
130 |
+
## Links
|
131 |
+
|
132 |
+
* Paper: Paper: [WhisperNER: Unified Open Named Entity and Speech Recognition](https://arxiv.org/abs/2409.08107).
|
133 |
+
* Model: https://huggingface.co/aiola/whisper-ner-v1
|
134 |
+
* Code: https://github.com/aiola-lab/whisper-ner
|
135 |
+
"""
|
136 |
+
)
|
137 |
+
|
138 |
+
with gr.Row() as row1:
|
139 |
+
with gr.Column() as col1:
|
140 |
+
audio_input = gr.Audio(label="Audio Example", type="filepath")
|
141 |
+
with gr.Column() as col2:
|
142 |
+
label_input = gr.Textbox(label="Entity Labels")
|
143 |
+
|
144 |
+
gr.Markdown("## Output")
|
145 |
+
|
146 |
+
with gr.Row() as row3:
|
147 |
+
transcript_output = gr.Textbox(label="Transcription and Entities")
|
148 |
+
|
149 |
+
with gr.Row() as row4:
|
150 |
+
highlighted_text_output = gr.HighlightedText(label="Predicted Highlighted Entities")
|
151 |
+
|
152 |
+
submit_btn = gr.Button("Submit")
|
153 |
+
examples = gr.Examples(
|
154 |
+
examples,
|
155 |
+
fn=transcribe_and_recognize_entities,
|
156 |
+
inputs=[audio_input, label_input],
|
157 |
+
outputs=[transcript_output, highlighted_text_output],
|
158 |
+
cache_examples=True,
|
159 |
+
run_on_click=True,
|
160 |
+
)
|
161 |
+
|
162 |
+
# Submitting
|
163 |
+
label_input.submit(
|
164 |
+
fn=transcribe_and_recognize_entities,
|
165 |
+
inputs=[audio_input, label_input],
|
166 |
+
outputs=[transcript_output, highlighted_text_output],
|
167 |
+
)
|
168 |
+
|
169 |
+
demo.launch()
|