Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,126 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
3 |
+
import librosa
|
4 |
+
import torch
|
5 |
+
import epitran
|
6 |
+
import re
|
7 |
+
import difflib
|
8 |
+
import editdistance
|
9 |
+
from jiwer import wer
|
10 |
+
import json
|
11 |
+
|
12 |
+
# Load model once at startup
|
13 |
+
model_name = "jonatasgrosman/wav2vec2-large-xlsr-53-arabic"
|
14 |
+
processor = Wav2Vec2Processor.from_pretrained(model_name)
|
15 |
+
model = Wav2Vec2ForCTC.from_pretrained(model_name)
|
16 |
+
epi = epitran.Epitran('ara-Arab')
|
17 |
+
|
18 |
+
def clean_phonemes(ipa):
|
19 |
+
"""Remove diacritics and length markers from phonemes"""
|
20 |
+
return re.sub(r'[\u064B-\u0652\u02D0]', '', ipa)
|
21 |
+
|
22 |
+
def analyze_phonemes(language, reference_text, audio_file):
|
23 |
+
# Convert reference text to phonemes
|
24 |
+
ref_phonemes = []
|
25 |
+
for word in reference_text.split():
|
26 |
+
ipa = epi.transliterate(word)
|
27 |
+
ipa_clean = clean_phonemes(ipa)
|
28 |
+
ref_phonemes.append(list(ipa_clean))
|
29 |
+
|
30 |
+
# Process audio file
|
31 |
+
audio, sr = librosa.load(audio_file.name, sr=16000)
|
32 |
+
input_values = processor(audio, sampling_rate=16000, return_tensors="pt").input_values
|
33 |
+
|
34 |
+
# Get transcription
|
35 |
+
with torch.no_grad():
|
36 |
+
logits = model(input_values).logits
|
37 |
+
pred_ids = torch.argmax(logits, dim=-1)
|
38 |
+
transcription = processor.batch_decode(pred_ids)[0].strip()
|
39 |
+
|
40 |
+
# Convert transcription to phonemes
|
41 |
+
obs_phonemes = []
|
42 |
+
for word in transcription.split():
|
43 |
+
ipa = epi.transliterate(word)
|
44 |
+
ipa_clean = clean_phonemes(ipa)
|
45 |
+
obs_phonemes.append(list(ipa_clean))
|
46 |
+
|
47 |
+
# Prepare results in JSON format
|
48 |
+
results = {
|
49 |
+
"reference_text": reference_text,
|
50 |
+
"transcription": transcription,
|
51 |
+
"word_alignment": [],
|
52 |
+
"metrics": {}
|
53 |
+
}
|
54 |
+
|
55 |
+
# Calculate metrics
|
56 |
+
total_phoneme_errors = 0
|
57 |
+
total_phoneme_length = 0
|
58 |
+
correct_words = 0
|
59 |
+
total_word_length = len(ref_phonemes)
|
60 |
+
|
61 |
+
# Word-by-word alignment
|
62 |
+
for i, (ref, obs) in enumerate(zip(ref_phonemes, obs_phonemes)):
|
63 |
+
ref_str = ''.join(ref)
|
64 |
+
obs_str = ''.join(obs)
|
65 |
+
edits = editdistance.eval(ref, obs)
|
66 |
+
acc = round((1 - edits / max(1, len(ref))) * 100, 2)
|
67 |
+
|
68 |
+
# Get error details
|
69 |
+
matcher = difflib.SequenceMatcher(None, ref, obs)
|
70 |
+
ops = matcher.get_opcodes()
|
71 |
+
error_details = []
|
72 |
+
for tag, i1, i2, j1, j2 in ops:
|
73 |
+
ref_seg = ''.join(ref[i1:i2]) or '-'
|
74 |
+
obs_seg = ''.join(obs[j1:j2]) or '-'
|
75 |
+
if tag != 'equal':
|
76 |
+
error_details.append({
|
77 |
+
"type": tag.upper(),
|
78 |
+
"reference": ref_seg,
|
79 |
+
"observed": obs_seg
|
80 |
+
})
|
81 |
+
|
82 |
+
results["word_alignment"].append({
|
83 |
+
"word_index": i,
|
84 |
+
"reference_phonemes": ref_str,
|
85 |
+
"observed_phonemes": obs_str,
|
86 |
+
"edit_distance": edits,
|
87 |
+
"accuracy": acc,
|
88 |
+
"is_correct": edits == 0,
|
89 |
+
"errors": error_details
|
90 |
+
})
|
91 |
+
|
92 |
+
total_phoneme_errors += edits
|
93 |
+
total_phoneme_length += len(ref)
|
94 |
+
correct_words += 1 if edits == 0 else 0
|
95 |
+
|
96 |
+
# Calculate metrics
|
97 |
+
phoneme_acc = round((1 - total_phoneme_errors / max(1, total_phoneme_length)) * 100, 2)
|
98 |
+
phoneme_er = round((total_phoneme_errors / max(1, total_phoneme_length)) * 100, 2)
|
99 |
+
word_acc = round((correct_words / max(1, total_word_length)) * 100, 2)
|
100 |
+
word_er = round(((total_word_length - correct_words) / max(1, total_word_length)) * 100, 2)
|
101 |
+
text_wer = round(wer(reference_text, transcription) * 100, 2)
|
102 |
+
|
103 |
+
results["metrics"] = {
|
104 |
+
"word_accuracy": word_acc,
|
105 |
+
"word_error_rate": word_er,
|
106 |
+
"phoneme_accuracy": phoneme_acc,
|
107 |
+
"phoneme_error_rate": phoneme_er,
|
108 |
+
"asr_word_error_rate": text_wer
|
109 |
+
}
|
110 |
+
|
111 |
+
return json.dumps(results, indent=2, ensure_ascii=False)
|
112 |
+
|
113 |
+
# Create Gradio interface
|
114 |
+
demo = gr.Interface(
|
115 |
+
fn=analyze_phonemes,
|
116 |
+
inputs=[
|
117 |
+
gr.Dropdown(["Arabic"], label="Language", value="Arabic"),
|
118 |
+
gr.Textbox(label="Reference Text", value="ููุจูุฃูููู ุขููุงุกู ุฑูุจููููู
ูุง ุชูููุฐููุจูุงูู"),
|
119 |
+
gr.File(label="Upload Audio File", type="file")
|
120 |
+
],
|
121 |
+
outputs=gr.JSON(label="Phoneme Alignment Results"),
|
122 |
+
title="Arabic Phoneme Alignment Analysis",
|
123 |
+
description="Compare audio pronunciation with reference text at phoneme level"
|
124 |
+
)
|
125 |
+
|
126 |
+
demo.launch()
|