sudhanm's picture
Update app.py
05566a8 verified
raw
history blame
25.5 kB
import gradio as gr
import random
import difflib
import re
import jiwer
import torch
import warnings
import contextlib
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq, pipeline
import librosa
import numpy as np
# Optional transliteration
try:
from indic_transliteration import sanscript
from indic_transliteration.sanscript import transliterate
INDIC_OK = True
except:
INDIC_OK = False
print("⚠️ indic_transliteration not available. Transliteration features disabled.")
# Optional HF Spaces GPU decorator
try:
import spaces
GPU_DECORATOR = spaces.GPU
except:
class _NoOp:
def __call__(self, f): return f
GPU_DECORATOR = _NoOp()
warnings.filterwarnings("ignore")
# ---------------- CONFIG ---------------- #
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DEVICE_INDEX = 0 if DEVICE == "cuda" else -1
DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32
amp_ctx = torch.cuda.amp.autocast if DEVICE == "cuda" else contextlib.nullcontext
print(f"🔧 Using device: {DEVICE}")
LANG_CODES = {
"English": "en",
"Tamil": "ta",
"Malayalam": "ml",
"Hindi": "hi"
}
# Primary: IndicWhisper
INDICWHISPER_MODEL = "parthiv11/indic_whisper_nodcil"
# Specialized models for better accuracy
SPECIALIZED_MODELS = {
"English": "openai/whisper-base.en",
"Tamil": "vasista22/whisper-tamil-large-v2",
"Malayalam": "thennal/whisper-medium-ml",
"Hindi": "openai/whisper-large-v2" # Using general model for Hindi
}
SCRIPT_PATTERNS = {
"Tamil": re.compile(r"[஀-௿]"),
"Malayalam": re.compile(r"[ഀ-ൿ]"),
"Hindi": re.compile(r"[ऀ-ॿ]"),
"English": re.compile(r"[A-Za-z]")
}
# Transliteration mappings
TRANSLITERATION_SCRIPTS = {
"Tamil": sanscript.TAMIL,
"Malayalam": sanscript.MALAYALAM,
"Hindi": sanscript.DEVANAGARI,
"English": None
}
SENTENCE_BANK = {
"English": [
"The sun sets over the horizon.",
"Learning languages is fun and rewarding.",
"I like to drink coffee in the morning.",
"Technology helps us connect with others.",
"Reading books expands our knowledge."
],
"Tamil": [
"இன்று நல்ல வானிலை உள்ளது.",
"நான் தமிழ் கற்றுக்கொண்டு இருக்கிறேன்.",
"எனக்கு புத்தகம் படிக்க விருப்பம்.",
"காலையில் காபி குடிக்க பிடிக்கும்.",
"நண்பர்களுடன் பேசுவது மகிழ்ச்சி."
],
"Malayalam": [
"എനിക്ക് മലയാളം വളരെ ഇഷ്ടമാണ്.",
"ഇന്ന് മഴപെയ്യുന്നു.",
"ഞാൻ പുസ്തകം വായിക്കുന്നു.",
"കാലയിൽ ചായ കുടിക്കാൻ ഇഷ്ടമാണ്.",
"സുഹൃത്തുക്കളോടു സംസാരിക്കുന്നത് സന്തോഷമാണ്."
],
"Hindi": [
"आज मौसम अच्छा है।",
"मुझे हिंदी बोलना पसंद है।",
"मैं किताब पढ़ रहा हूँ।",
"सुबह चाय पीना अच्छा लगता है।",
"दोस्तों के साथ बात करना खुशी देता है।"
]
}
# Model cache
primary_pipeline = None
specialized_models = {}
# ---------------- HELPERS ---------------- #
def get_random_sentence(language_choice):
return random.choice(SENTENCE_BANK[language_choice])
def is_correct_script(text, lang_name):
"""Check if text contains the expected script for the language"""
if not text.strip():
return False
pattern = SCRIPT_PATTERNS.get(lang_name)
if not pattern:
return True
return bool(pattern.search(text))
def transliterate_text(text, lang_choice, to_romanized=True):
"""Transliterate text to/from romanized form"""
if not INDIC_OK or not text.strip():
return text
source_script = TRANSLITERATION_SCRIPTS.get(lang_choice)
if not source_script:
return text
try:
if to_romanized:
# Convert to Harvard-Kyoto (romanized)
return transliterate(text, source_script, sanscript.HK)
else:
# Convert from romanized to native script (if needed)
return transliterate(text, sanscript.HK, source_script)
except Exception as e:
print(f"⚠️ Transliteration failed: {e}")
return text
def preprocess_audio(audio_path, target_sr=16000):
"""Enhanced audio preprocessing"""
try:
audio, sr = librosa.load(audio_path, sr=target_sr, mono=True)
if audio is None or len(audio) == 0:
return None, None
# Normalize audio
audio = audio.astype(np.float32)
max_val = np.max(np.abs(audio))
if max_val > 0:
audio = audio / max_val
# Trim silence
audio, _ = librosa.effects.trim(audio, top_db=20)
# Check minimum length (0.1 seconds)
if len(audio) < int(target_sr * 0.1):
return None, None
return audio, target_sr
except Exception as e:
print(f"⚠️ Audio preprocessing failed: {e}")
return None, None
# ---------------- MODEL LOADERS ---------------- #
@GPU_DECORATOR
def load_primary_model():
"""Load the primary IndicWhisper model"""
global primary_pipeline
if primary_pipeline is not None:
return primary_pipeline
try:
print(f"🔄 Loading primary model: {INDICWHISPER_MODEL}")
# Try direct loading first
primary_pipeline = pipeline(
"automatic-speech-recognition",
model=INDICWHISPER_MODEL,
device=DEVICE_INDEX,
torch_dtype=DTYPE,
trust_remote_code=True
)
print("✅ Primary model loaded successfully!")
return primary_pipeline
except Exception as e:
print(f"⚠️ Primary model failed, using fallback: {e}")
# Fallback to base Whisper
primary_pipeline = pipeline(
"automatic-speech-recognition",
model="openai/whisper-large-v2",
device=DEVICE_INDEX,
torch_dtype=DTYPE
)
print("✅ Fallback model loaded!")
return primary_pipeline
@GPU_DECORATOR
def load_specialized_model(language):
"""Load specialized model for specific language"""
if language in specialized_models:
return specialized_models[language]
model_name = SPECIALIZED_MODELS[language]
print(f"🔄 Loading specialized {language} model: {model_name}")
try:
processor = AutoProcessor.from_pretrained(model_name)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
model_name,
torch_dtype=DTYPE,
device_map="auto" if DEVICE == "cuda" else None
).to(DEVICE)
specialized_models[language] = {
"processor": processor,
"model": model
}
print(f"✅ Specialized {language} model loaded!")
return specialized_models[language]
except Exception as e:
print(f"❌ Failed to load specialized {language} model: {e}")
return None
# ---------------- TRANSCRIPTION ---------------- #
@GPU_DECORATOR
def transcribe_with_primary(audio_path, language):
"""Transcribe using primary IndicWhisper model"""
try:
pipeline_model = load_primary_model()
lang_code = LANG_CODES[language]
# Set language forcing if possible
try:
if hasattr(pipeline_model, "model") and hasattr(pipeline_model, "tokenizer"):
forced_ids = pipeline_model.tokenizer.get_decoder_prompt_ids(
language=lang_code,
task="transcribe"
)
if forced_ids:
pipeline_model.model.config.forced_decoder_ids = forced_ids
except Exception as e:
print(f"⚠️ Language forcing failed: {e}")
with amp_ctx():
result = pipeline_model(audio_path)
if isinstance(result, dict):
return result.get("text", "").strip()
return str(result).strip()
except Exception as e:
return f"Primary transcription error: {str(e)}"
@GPU_DECORATOR
def transcribe_with_specialized(audio_path, language):
"""Transcribe using specialized model"""
try:
model_components = load_specialized_model(language)
if not model_components:
return "Specialized model not available"
# Preprocess audio
audio, sr = preprocess_audio(audio_path)
if audio is None:
return "Audio preprocessing failed"
# Process with specialized model
inputs = model_components["processor"](
audio,
sampling_rate=sr,
return_tensors="pt"
)
input_features = inputs.input_features.to(DEVICE)
# Generation parameters
gen_kwargs = {
"inputs": input_features,
"max_length": 200,
"num_beams": 3,
"do_sample": False
}
# Language forcing for non-English
if language != "English":
try:
forced_ids = model_components["processor"].tokenizer.get_decoder_prompt_ids(
language=LANG_CODES[language],
task="transcribe"
)
if forced_ids:
gen_kwargs["forced_decoder_ids"] = forced_ids
except Exception as e:
print(f"⚠️ Specialized language forcing failed: {e}")
# Generate transcription
with torch.no_grad(), amp_ctx():
generated_ids = model_components["model"].generate(**gen_kwargs)
# Decode result
transcription = model_components["processor"].batch_decode(
generated_ids,
skip_special_tokens=True
)[0]
return transcription.strip()
except Exception as e:
return f"Specialized transcription error: {str(e)}"
# ---------------- ANALYSIS ---------------- #
def compute_metrics(reference, hypothesis):
"""Compute WER and CER with error handling"""
try:
# Clean up texts
ref_clean = reference.strip()
hyp_clean = hypothesis.strip()
if not ref_clean or not hyp_clean:
return 1.0, 1.0
# Compute WER and CER
wer = jiwer.wer(ref_clean, hyp_clean)
cer = jiwer.cer(ref_clean, hyp_clean)
return wer, cer
except Exception as e:
print(f"⚠️ Metric computation failed: {e}")
return 1.0, 1.0
def get_pronunciation_score(wer, cer):
"""Convert error rates to intuitive scores and feedback"""
# Weighted combination (WER is more important)
combined_error = (wer * 0.7) + (cer * 0.3)
accuracy = 1 - combined_error
if accuracy >= 0.95:
return "🏆 Perfect!", "Outstanding pronunciation! Native-like accuracy.", "#d4edda"
elif accuracy >= 0.85:
return "🎉 Excellent!", "Very good pronunciation with minor variations.", "#d1ecf1"
elif accuracy >= 0.70:
return "👍 Good!", "Good pronunciation, practice specific sounds.", "#fff3cd"
elif accuracy >= 0.50:
return "📚 Needs Practice", "Focus on clearer pronunciation and rhythm.", "#f8d7da"
else:
return "💪 Keep Trying!", "Break down into smaller parts and practice slowly.", "#f5c6cb"
def create_detailed_comparison(intended, actual, lang_choice):
"""Create detailed side-by-side comparison with transliteration"""
# Original scripts
intended_orig = intended.strip()
actual_orig = actual.strip()
# Transliterations
intended_translit = transliterate_text(intended_orig, lang_choice, to_romanized=True)
actual_translit = transliterate_text(actual_orig, lang_choice, to_romanized=True)
# Word-level highlighting
word_diff_orig = highlight_word_differences(intended_orig, actual_orig)
word_diff_translit = highlight_word_differences(intended_translit, actual_translit)
# Character-level highlighting
char_diff_orig = highlight_char_differences(intended_orig, actual_orig)
char_diff_translit = highlight_char_differences(intended_translit, actual_translit)
return {
"intended_orig": intended_orig,
"actual_orig": actual_orig,
"intended_translit": intended_translit,
"actual_translit": actual_translit,
"word_diff_orig": word_diff_orig,
"word_diff_translit": word_diff_translit,
"char_diff_orig": char_diff_orig,
"char_diff_translit": char_diff_translit
}
def highlight_word_differences(reference, hypothesis):
"""Highlight word-level differences with colors"""
ref_words = reference.split()
hyp_words = hypothesis.split()
sm = difflib.SequenceMatcher(None, ref_words, hyp_words)
html_output = []
for tag, i1, i2, j1, j2 in sm.get_opcodes():
if tag == 'equal':
# Correct words - green background
html_output.extend([
f"<span style='background-color:#d4edda; color:#155724; padding:2px 4px; margin:1px; border-radius:3px'>{word}</span>"
for word in ref_words[i1:i2]
])
elif tag == 'replace':
# Wrong words - red background for reference, orange for hypothesis
html_output.extend([
f"<span style='background-color:#f8d7da; color:#721c24; padding:2px 4px; margin:1px; border-radius:3px; text-decoration:line-through'>{word}</span>"
for word in ref_words[i1:i2]
])
html_output.extend([
f"<span style='background-color:#fff3cd; color:#856404; padding:2px 4px; margin:1px; border-radius:3px'>→{word}</span>"
for word in hyp_words[j1:j2]
])
elif tag == 'delete':
# Missing words - red background
html_output.extend([
f"<span style='background-color:#f8d7da; color:#721c24; padding:2px 4px; margin:1px; border-radius:3px; text-decoration:line-through'>{word}</span>"
for word in ref_words[i1:i2]
])
elif tag == 'insert':
# Extra words - orange background
html_output.extend([
f"<span style='background-color:#fff3cd; color:#856404; padding:2px 4px; margin:1px; border-radius:3px'>+{word}</span>"
for word in hyp_words[j1:j2]
])
return " ".join(html_output)
def highlight_char_differences(reference, hypothesis):
"""Highlight character-level differences"""
sm = difflib.SequenceMatcher(None, list(reference), list(hypothesis))
html_output = []
for tag, i1, i2, j1, j2 in sm.get_opcodes():
if tag == 'equal':
# Correct characters - green
html_output.extend([
f"<span style='color:#28a745'>{char}</span>"
for char in reference[i1:i2]
])
elif tag in ('replace', 'delete'):
# Wrong/missing characters - red with underline
html_output.extend([
f"<span style='color:#dc3545; text-decoration:underline; font-weight:bold'>{char}</span>"
for char in reference[i1:i2]
])
elif tag == 'insert':
# Extra characters - orange
html_output.extend([
f"<span style='color:#fd7e14; font-weight:bold'>{char}</span>"
for char in hypothesis[j1:j2]
])
return "".join(html_output)
def analyze_pronunciation_errors(intended, actual, lang_choice):
"""Provide specific feedback about pronunciation errors"""
comparison = create_detailed_comparison(intended, actual, lang_choice)
# Analyze error patterns
intended_words = intended.split()
actual_words = actual.split()
error_analysis = []
# Length difference analysis
if len(actual_words) < len(intended_words):
missing_count = len(intended_words) - len(actual_words)
error_analysis.append(f"🔍 You missed {missing_count} word(s). Try speaking more slowly.")
elif len(actual_words) > len(intended_words):
extra_count = len(actual_words) - len(intended_words)
error_analysis.append(f"🔍 You added {extra_count} extra word(s). Focus on the exact sentence.")
# Script verification
if not is_correct_script(actual, lang_choice):
error_analysis.append(f"⚠️ The transcription doesn't contain {lang_choice} script. Check your pronunciation.")
# WER/CER based feedback
wer, cer = compute_metrics(intended, actual)
if wer > 0.5:
error_analysis.append("🎯 Focus on pronouncing each word clearly and separately.")
elif wer > 0.3:
error_analysis.append("🎯 Good overall, but some words need clearer pronunciation.")
if cer > 0.3:
error_analysis.append("🔤 Pay attention to individual sounds and syllables.")
return error_analysis, comparison
# ---------------- MAIN FUNCTION ---------------- #
@GPU_DECORATOR
def compare_pronunciation(audio, language_choice, intended_sentence):
"""Main function to analyze pronunciation"""
if audio is None:
return ("❌ Please record audio first", "", "", "", "", "", "", "", "", "", "")
if not intended_sentence.strip():
return ("❌ Please generate a sentence first", "", "", "", "", "", "", "", "", "", "")
print(f"🔍 Analyzing pronunciation for {language_choice}...")
# Get transcriptions from both models
primary_result = transcribe_with_primary(audio, language_choice)
specialized_result = transcribe_with_specialized(audio, language_choice)
# Choose best result (prefer specialized if successful)
if not specialized_result.startswith("Specialized") and specialized_result.strip():
best_transcription = specialized_result
best_source = "Specialized Model"
elif not primary_result.startswith("Primary") and primary_result.strip():
best_transcription = primary_result
best_source = "Primary Model"
else:
return (
f"❌ Both models failed:\nPrimary: {primary_result}\nSpecialized: {specialized_result}",
"", "", "", "", "", "", "", "", "", ""
)
# Analyze pronunciation
error_analysis, comparison = analyze_pronunciation_errors(
intended_sentence, best_transcription, language_choice
)
# Compute metrics
wer, cer = compute_metrics(intended_sentence, best_transcription)
score, feedback, color = get_pronunciation_score(wer, cer)
# Create status message
status_msg = f"""✅ Analysis Complete!
{score}
{feedback}
🤖 Best result from: {best_source}
📊 Word Accuracy: {(1-wer)*100:.1f}%
📈 Character Accuracy: {(1-cer)*100:.1f}%
🔍 Analysis:
""" + "\n".join(error_analysis)
return (
status_msg,
primary_result,
specialized_result,
f"{wer:.3f} ({(1-wer)*100:.1f}%)",
f"{cer:.3f} ({(1-cer)*100:.1f}%)",
comparison["intended_orig"],
comparison["actual_orig"],
comparison["intended_translit"],
comparison["actual_translit"],
comparison["word_diff_orig"],
comparison["char_diff_orig"]
)
# ---------------- UI ---------------- #
def create_interface():
with gr.Blocks(title="Enhanced Pronunciation Comparator", theme=gr.themes.Soft()) as demo:
gr.Markdown("""
# 🎙️ Enhanced Pronunciation Comparator
**Perfect your pronunciation in English, Tamil, Malayalam, and Hindi!**
This tool uses specialized AI models to give you detailed feedback on your pronunciation,
including transliteration to help you understand exactly where you need improvement.
### How to use:
1. 🌐 Select your target language
2. 🎲 Generate a practice sentence
3. 🎤 Record yourself saying the sentence clearly
4. 🔍 Get detailed pronunciation analysis with transliteration
""")
with gr.Row():
with gr.Column(scale=2):
language_dropdown = gr.Dropdown(
choices=list(LANG_CODES.keys()),
value="Tamil",
label="🌐 Select Language"
)
with gr.Column(scale=1):
generate_btn = gr.Button("🎲 Generate Practice Sentence", variant="primary")
intended_textbox = gr.Textbox(
label="📝 Practice Sentence",
interactive=False,
lines=2,
placeholder="Click 'Generate Practice Sentence' to get started..."
)
audio_input = gr.Audio(
sources=["microphone", "upload"],
type="filepath",
label="🎤 Record Your Pronunciation"
)
analyze_btn = gr.Button("🔍 Analyze Pronunciation", variant="secondary", size="lg")
with gr.Row():
status_output = gr.Textbox(
label="📊 Analysis Results",
interactive=False,
lines=8
)
with gr.Accordion("🤖 Model Outputs", open=False):
with gr.Row():
primary_output = gr.Textbox(label="Primary Model (IndicWhisper)", interactive=False)
specialized_output = gr.Textbox(label="Specialized Model", interactive=False)
with gr.Accordion("📈 Detailed Metrics", open=False):
with gr.Row():
wer_output = gr.Textbox(label="Word Error Rate", interactive=False)
cer_output = gr.Textbox(label="Character Error Rate", interactive=False)
gr.Markdown("### 🔍 Detailed Comparison")
with gr.Row():
with gr.Column():
gr.Markdown("#### 📝 Original Script")
intended_orig = gr.Textbox(label="🎯 Target Text", interactive=False)
actual_orig = gr.Textbox(label="🗣️ What You Said", interactive=False)
with gr.Column():
gr.Markdown("#### 🔤 Romanized (Transliterated)")
intended_translit = gr.Textbox(label="🎯 Target (Romanized)", interactive=False)
actual_translit = gr.Textbox(label="🗣️ What You Said (Romanized)", interactive=False)
gr.Markdown("### 🎨 Visual Comparison")
gr.Markdown("**Green** = Correct, **Red** = Wrong/Missing, **Orange** = Added/Substituted")
word_diff_html = gr.HTML(label="🔤 Word-by-Word Comparison")
char_diff_html = gr.HTML(label="🔍 Character-by-Character Analysis")
# Event handlers
generate_btn.click(
fn=get_random_sentence,
inputs=[language_dropdown],
outputs=[intended_textbox]
)
analyze_btn.click(
fn=compare_pronunciation,
inputs=[audio_input, language_dropdown, intended_textbox],
outputs=[
status_output, primary_output, specialized_output,
wer_output, cer_output, intended_orig, actual_orig,
intended_translit, actual_translit, word_diff_html, char_diff_html
]
)
language_dropdown.change(
fn=get_random_sentence,
inputs=[language_dropdown],
outputs=[intended_textbox]
)
gr.Markdown("""
### 📚 Pro Tips for Better Pronunciation:
- **Speak slowly and clearly** - Don't rush through the sentence
- **Pronounce each syllable** - Break down complex words
- **Check the romanized version** - Use it to understand correct pronunciation
- **Practice repeatedly** - Use the same sentence multiple times to track improvement
- **Focus on problem areas** - Pay attention to red-highlighted parts
- **Record in a quiet environment** - Minimize background noise
### 🎯 Understanding the Feedback:
- **Green highlights** = Perfect pronunciation ✅
- **Red highlights** = Missing or mispronounced ❌
- **Orange highlights** = Added or substituted 🔄
- **Transliteration** = Helps you see pronunciation patterns
- **Error rates** = Lower is better (0% = perfect)
""")
return demo
# ---------------- LAUNCH ---------------- #
if __name__ == "__main__":
print("🚀 Starting Enhanced Pronunciation Comparator...")
demo = create_interface()
demo.launch(
server_name="0.0.0.0",
server_port=7860,
share=True,
show_error=True
)