wav2vec2 / app.py
101Frost's picture
Update app.py
c64435a verified
import gradio as gr
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import librosa
import torch
import epitran
import re
import difflib
import editdistance
from jiwer import wer
import json
import string
import eng_to_ipa as ipa
# Load both models at startup
MODELS = {
"Arabic": {
"processor": Wav2Vec2Processor.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-arabic"),
"model": Wav2Vec2ForCTC.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-arabic"),
"epitran": epitran.Epitran("ara-Arab")
},
"English": {
"processor": Wav2Vec2Processor.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-english"),
"model": Wav2Vec2ForCTC.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-english"),
"epitran": epitran.Epitran("eng-Latn")
}
}
# Suppress the warning about newly initialized weights
for lang in MODELS.values():
lang["model"].config.ctc_loss_reduction = "mean"
def clean_phonemes(ipa_text):
"""Remove diacritics and length markers from phonemes"""
return re.sub(r'[\u064B-\u0652\u02D0]', '', ipa_text)
def safe_transliterate_arabic(epi, word):
try:
word = word.strip()
ipa = epi.transliterate(word)
if not ipa.strip():
raise ValueError("Empty IPA string")
return clean_phonemes(ipa)
except Exception as e:
print(f"[Warning] Arabic transliteration failed for '{word}': {e}")
return ""
def transliterate_english(word):
try:
word = word.lower().translate(str.maketrans('', '', string.punctuation))
ipa_text = ipa.convert(word)
return clean_phonemes(ipa_text)
except Exception as e:
print(f"[Warning] English IPA conversion failed for '{word}': {e}")
return ""
def analyze_phonemes(language, reference_text, audio_file):
# Get the appropriate model components
lang_models = MODELS[language]
processor = lang_models["processor"]
model = lang_models["model"]
epi = lang_models["epitran"]
if language == "Arabic":
transliterate_fn = lambda word: safe_transliterate_arabic(epi, word)
else:
transliterate_fn = transliterate_english
# Convert reference text to phonemes
ref_phonemes = []
for word in reference_text.split():
ipa_clean = transliterate_fn(word)
ref_phonemes.append(list(ipa_clean))
# Process audio file
audio, sr = librosa.load(audio_file, sr=16000)
input_values = processor(audio, sampling_rate=16000, return_tensors="pt").input_values
# Get transcription
with torch.no_grad():
logits = model(input_values).logits
pred_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(pred_ids)[0].strip()
# Convert transcription to phonemes
obs_phonemes = []
for word in transcription.split():
ipa_clean = transliterate_fn(word)
obs_phonemes.append(list(ipa_clean))
# Prepare results in JSON format
results = {
"language": language,
"reference_text": reference_text,
"transcription": transcription,
"word_alignment": [],
"metrics": {}
}
# Calculate metrics
total_phoneme_errors = 0
total_phoneme_length = 0
correct_words = 0
total_word_length = len(ref_phonemes)
# Word-by-word alignment
for i, (ref, obs) in enumerate(zip(ref_phonemes, obs_phonemes)):
ref_str = ''.join(ref)
obs_str = ''.join(obs)
edits = editdistance.eval(ref, obs)
acc = round((1 - edits / max(1, len(ref))) * 100, 2)
# Get error details
matcher = difflib.SequenceMatcher(None, ref, obs)
ops = matcher.get_opcodes()
error_details = []
for tag, i1, i2, j1, j2 in ops:
ref_seg = ''.join(ref[i1:i2]) or '-'
obs_seg = ''.join(obs[j1:j2]) or '-'
if tag != 'equal':
error_details.append({
"type": tag.upper(),
"reference": ref_seg,
"observed": obs_seg
})
results["word_alignment"].append({
"word_index": i,
"reference_phonemes": ref_str,
"observed_phonemes": obs_str,
"edit_distance": edits,
"accuracy": acc,
"is_correct": edits == 0,
"errors": error_details
})
total_phoneme_errors += edits
total_phoneme_length += len(ref)
correct_words += 1 if edits == 0 else 0
# Final metrics
phoneme_acc = round((1 - total_phoneme_errors / max(1, total_phoneme_length)) * 100, 2)
phoneme_er = round((total_phoneme_errors / max(1, total_phoneme_length)) * 100, 2)
word_acc = round((correct_words / max(1, total_word_length)) * 100, 2)
word_er = round(((total_word_length - correct_words) / max(1, total_word_length)) * 100, 2)
text_wer = round(wer(reference_text, transcription) * 100, 2)
results["metrics"] = {
"word_accuracy": word_acc,
"word_error_rate": word_er,
"phoneme_accuracy": phoneme_acc,
"phoneme_error_rate": phoneme_er,
"asr_word_error_rate": text_wer
}
return json.dumps(results, indent=2, ensure_ascii=False)
# Create Gradio interface with language-specific default text
def get_default_text(language):
return {
"Arabic": "ููŽุจูุฃูŽูŠู‘ู ุขู„ูŽุงุกู ุฑูŽุจู‘ููƒูู…ูŽุง ุชููƒูŽุฐู‘ูุจูŽุงู†ู",
"English": "The quick brown fox jumps over the lazy dog"
}.get(language, "")
with gr.Blocks() as demo:
gr.Markdown("# Multilingual Phoneme Alignment Analysis")
gr.Markdown("Compare audio pronunciation with reference text at phoneme level")
with gr.Row():
language = gr.Dropdown(
["Arabic", "English"],
label="Language",
value="Arabic"
)
reference_text = gr.Textbox(
label="Reference Text",
value=get_default_text("Arabic")
)
audio_input = gr.Audio(label="Upload Audio File", type="filepath")
submit_btn = gr.Button("Analyze")
output = gr.JSON(label="Phoneme Alignment Results")
language.change(
fn=get_default_text,
inputs=language,
outputs=reference_text
)
submit_btn.click(
fn=analyze_phonemes,
inputs=[language, reference_text, audio_input],
outputs=output
)
demo.launch()