import os
import io
import gradio as gr
import torch
import numpy as np
import re
import pronouncing
import functools
from transformers import (
AutoModelForAudioClassification,
AutoFeatureExtractor,
AutoTokenizer,
pipeline,
AutoModelForCausalLM,
BitsAndBytesConfig
)
from huggingface_hub import login
from utils import (
load_audio,
extract_audio_duration,
extract_mfcc_features,
format_genre_results,
ensure_cuda_availability
)
from emotionanalysis import MusicAnalyzer
import librosa
from beat_analysis import BeatAnalyzer # Import the BeatAnalyzer class
# Initialize beat analyzer
beat_analyzer = BeatAnalyzer()
# Login to Hugging Face Hub if token is provided
if "HF_TOKEN" in os.environ:
login(token=os.environ["HF_TOKEN"])
# Constants
GENRE_MODEL_NAME = "dima806/music_genres_classification"
MUSIC_DETECTION_MODEL = "MIT/ast-finetuned-audioset-10-10-0.4593"
LLM_MODEL_NAME = "Qwen/Qwen3-32B"
SAMPLE_RATE = 22050 # Standard sample rate for audio processing
# Check CUDA availability (for informational purposes)
CUDA_AVAILABLE = ensure_cuda_availability()
# Load models at initialization time
print("Loading genre classification model...")
try:
genre_feature_extractor = AutoFeatureExtractor.from_pretrained(GENRE_MODEL_NAME)
genre_model = AutoModelForAudioClassification.from_pretrained(
GENRE_MODEL_NAME,
device_map="auto" if CUDA_AVAILABLE else None
)
# Create a convenience wrapper function with the same interface as before
def get_genre_model():
return genre_model, genre_feature_extractor
except Exception as e:
print(f"Error loading genre model: {str(e)}")
genre_model = None
genre_feature_extractor = None
# Load LLM and tokenizer at initialization time
print("Loading Qwen LLM model with 4-bit quantization...")
try:
# Configure 4-bit quantization for better performance
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True
)
llm_tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL_NAME)
llm_model = AutoModelForCausalLM.from_pretrained(
LLM_MODEL_NAME,
quantization_config=quantization_config,
device_map="auto",
trust_remote_code=True,
torch_dtype=torch.float16,
use_cache=True
)
except Exception as e:
print(f"Error loading LLM model: {str(e)}")
llm_tokenizer = None
llm_model = None
# Create music analyzer instance
music_analyzer = MusicAnalyzer()
# Process uploaded audio file
def process_audio(audio_file):
if audio_file is None:
return "No audio file provided", None, None, None, None, None, None, None
try:
# Load and analyze audio
y, sr = load_audio(audio_file, sr=SAMPLE_RATE)
# Basic audio information
duration = extract_audio_duration(y, sr)
# Analyze music with MusicAnalyzer
music_analysis = music_analyzer.analyze_music(audio_file)
# Extract key information
tempo = music_analysis["rhythm_analysis"]["tempo"]
emotion = music_analysis["emotion_analysis"]["primary_emotion"]
theme = music_analysis["theme_analysis"]["primary_theme"]
# Use genre classification directly instead of pipeline
if genre_model is not None and genre_feature_extractor is not None:
# Resample audio to 16000 Hz for the genre model
y_16k = librosa.resample(y, orig_sr=sr, target_sr=16000)
# Extract features
inputs = genre_feature_extractor(
y_16k,
sampling_rate=16000,
return_tensors="pt"
).to(genre_model.device)
# Classify genre
with torch.no_grad():
outputs = genre_model(**inputs)
logits = outputs.logits
probs = torch.nn.functional.softmax(logits, dim=-1)
# Get top genres
values, indices = torch.topk(probs[0], k=5)
top_genres = [(genre_model.config.id2label[idx.item()], val.item()) for val, idx in zip(values, indices)]
else:
# Fallback if model loading failed
top_genres = [("Unknown", 1.0)]
# Format genre results for display
genre_results_text = format_genre_results(top_genres)
primary_genre = top_genres[0][0]
# Override time signature for pop and disco genres to always be 4/4
if any(genre.lower() in primary_genre.lower() for genre in ['pop', 'disco']):
music_analysis["rhythm_analysis"]["estimated_time_signature"] = "4/4"
time_signature = "4/4"
else:
# Use detected time signature for other genres
time_signature = music_analysis["rhythm_analysis"]["estimated_time_signature"]
# Ensure time signature is one of the supported ones (4/4, 3/4, 6/8)
if time_signature not in ["4/4", "3/4", "6/8"]:
time_signature = "4/4" # Default to 4/4 if unsupported
music_analysis["rhythm_analysis"]["estimated_time_signature"] = time_signature
# Analyze beat patterns and create lyrics template using the time signature
beat_analysis = beat_analyzer.analyze_beat_pattern(audio_file, time_signature=time_signature)
lyric_templates = beat_analyzer.create_lyric_template(beat_analysis)
# Store these in the music_analysis dict for use in lyrics generation
music_analysis["beat_analysis"] = beat_analysis
music_analysis["lyric_templates"] = lyric_templates
# Prepare analysis summary
analysis_summary = f"""
### Music Analysis Results
**Duration:** {duration:.2f} seconds
**Tempo:** {tempo:.1f} BPM
**Time Signature:** {time_signature}
**Key:** {music_analysis["tonal_analysis"]["key"]} {music_analysis["tonal_analysis"]["mode"]}
**Primary Emotion:** {emotion}
**Primary Theme:** {theme}
**Top Genre:** {primary_genre}
{genre_results_text}
"""
# Add beat analysis summary
if lyric_templates:
analysis_summary += f"""
### Beat Analysis
**Total Phrases:** {len(lyric_templates)}
**Average Beats Per Phrase:** {np.mean([t['num_beats'] for t in lyric_templates]):.1f}
**Beat Pattern Examples:**
- Phrase 1: {lyric_templates[0]['stress_pattern'] if lyric_templates else 'N/A'}
- Phrase 2: {lyric_templates[1]['stress_pattern'] if len(lyric_templates) > 1 else 'N/A'}
"""
# Check if genre is supported for lyrics generation
# Use the supported_genres list from BeatAnalyzer
genre_supported = any(genre.lower() in primary_genre.lower() for genre in beat_analyzer.supported_genres)
# Generate lyrics only for supported genres
if genre_supported:
lyrics = generate_lyrics(music_analysis, primary_genre, duration)
beat_match_analysis = analyze_lyrics_rhythm_match(lyrics, lyric_templates, primary_genre)
else:
supported_genres_str = ", ".join([genre.capitalize() for genre in beat_analyzer.supported_genres])
lyrics = f"Lyrics generation is only supported for the following genres: {supported_genres_str}.\n\nDetected genre '{primary_genre}' doesn't have strong syllable-to-beat patterns required for our lyric generation algorithm."
beat_match_analysis = "Lyrics generation not available for this genre."
return analysis_summary, lyrics, tempo, time_signature, emotion, theme, primary_genre, beat_match_analysis
except Exception as e:
error_msg = f"Error processing audio: {str(e)}"
print(error_msg)
return error_msg, None, None, None, None, None, None, None
def generate_lyrics(music_analysis, genre, duration):
try:
# Extract meaningful information for context
tempo = music_analysis["rhythm_analysis"]["tempo"]
key = music_analysis["tonal_analysis"]["key"]
mode = music_analysis["tonal_analysis"]["mode"]
emotion = music_analysis["emotion_analysis"]["primary_emotion"]
theme = music_analysis["theme_analysis"]["primary_theme"]
# Get beat analysis and templates
lyric_templates = music_analysis.get("lyric_templates", [])
# Define num_phrases here to ensure it's available in all code paths
num_phrases = len(lyric_templates) if lyric_templates else 4
# Verify LLM is loaded
if llm_model is None or llm_tokenizer is None:
return "Error: LLM model not properly loaded"
# If no templates, fall back to original method
if not lyric_templates:
# Simplified prompt
prompt = f"""Write song lyrics for a {genre} song in {key} {mode} with tempo {tempo} BPM. The emotion is {emotion} and theme is {theme}.
ONLY WRITE THE ACTUAL LYRICS. NO EXPLANATIONS OR META-TEXT.
"""
else:
# Calculate the typical syllable range for this genre
if num_phrases > 0:
# Get max syllables per line from templates
max_syllables = max([t.get('max_expected', 7) for t in lyric_templates]) if lyric_templates[0].get('max_expected') else 7
min_syllables = min([t.get('min_expected', 2) for t in lyric_templates]) if lyric_templates[0].get('min_expected') else 2
avg_syllables = (min_syllables + max_syllables) // 2
else:
min_syllables = 2
max_syllables = 7
avg_syllables = 4
# Create random examples based on the song's theme and emotion
# to avoid the LLM copying our examples directly
example_themes = [
{"emotion": "love", "fragments": ["I see your face", "across the room", "my heart beats fast", "can't look away"]},
{"emotion": "sadness", "fragments": ["tears fall like rain", "on empty streets", "memories fade", "into the dark"]},
{"emotion": "nostalgia", "fragments": ["old photographs", "dusty and worn", "remind me of when", "we were young"]},
{"emotion": "hope", "fragments": ["dawn breaks through clouds", "new day begins", "darkness recedes", "light fills my soul"]},
{"emotion": "longing", "fragments": ["miles apart now", "under same stars", "thinking of you", "across the distance"]}
]
# Select a theme that doesn't match the song's emotion to avoid copying
selected_themes = [t for t in example_themes if t["emotion"].lower() != emotion.lower()]
if not selected_themes:
selected_themes = example_themes
import random
example_theme = random.choice(selected_themes)
example_fragments = example_theme["fragments"]
random.shuffle(example_fragments) # Randomize order
# Create example 1 - grammatical connection with conjunction
ex1_line1 = example_fragments[0] if len(example_fragments) > 0 else "The morning sun"
ex1_line2 = example_fragments[1] if len(example_fragments) > 1 else "breaks through clouds"
ex1_line3 = example_fragments[2] if len(example_fragments) > 2 else "as birds begin"
ex1_line4 = example_fragments[3] if len(example_fragments) > 3 else "their dawn chorus"
# Create example 2 - prepositional connection
ex2_fragments = [
"She walks alone",
"through crowded streets",
"with memories",
"of better days"
]
random.shuffle(ex2_fragments)
# Create a more direct prompt with examples and specific syllable count guidance
prompt = f"""Write song lyrics for a {genre} song in {key} {mode} with tempo {tempo} BPM.
PRIMARY THEME: {theme}
EMOTION: {emotion}
I need EXACTLY {num_phrases} lines of lyrics with these STRICT requirements:
CRITICAL INSTRUCTIONS:
1. EXTREMELY SHORT LINES: Each line MUST be between {min_syllables}-{max_syllables} syllables MAXIMUM
2. ENFORCE BREVITY: NO exceptions to the syllable limit - not a single line should exceed {max_syllables} syllables
3. FRAGMENT STYLE: Use sentence fragments and short phrases instead of complete sentences
4. CONNECTED THOUGHTS: Use prepositions and conjunctions at the start of lines to connect ideas
5. SIMPLE WORDS: Choose one or two-syllable words whenever possible
6. CONCRETE IMAGERY: Use specific, tangible details rather than abstract concepts
7. NO CLICHÉS: Avoid common phrases like "time slips away" or "memories fade"
8. ONE THOUGHT PER LINE: Express just one simple idea in each line
FORMAT:
- Write exactly {num_phrases} short text lines
- No annotations, explanations, or line numbers
- Do not count syllables in the output
IMPORTANT: If you can't express an idea in {max_syllables} or fewer syllables, break it across two lines or choose a simpler way to express it.
===== EXAMPLES OF CORRECT LENGTH =====
Example 1 (short fragments connected by flow):
Cold tea cup (3 syllables)
on windowsill (3 syllables)
cat watches rain (3 syllables)
through foggy glass (3 syllables)
Example 2 (prepositional connections):
Keys dropped here (3 syllables)
by the front door (3 syllables)
where shoes pile up (3 syllables)
since you moved in (3 syllables)
DO NOT copy my examples. Create ENTIRELY NEW lyrics about {theme} with {emotion} feeling.
REMEMBER: NO LINE SHOULD EXCEED {max_syllables} SYLLABLES - this is the most important rule!
"""
# Generate lyrics using the LLM model
messages = [
{"role": "user", "content": prompt}
]
# Apply chat template
text = llm_tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
# Tokenize and move to model device
model_inputs = llm_tokenizer([text], return_tensors="pt").to(llm_model.device)
# Generate with optimized parameters
generated_ids = llm_model.generate(
**model_inputs,
max_new_tokens=1024,
do_sample=True,
temperature=0.7,
top_p=0.9,
repetition_penalty=1.2,
pad_token_id=llm_tokenizer.eos_token_id
)
# Decode the output
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
lyrics = llm_tokenizer.decode(output_ids, skip_special_tokens=True).strip()
# ULTRA AGGRESSIVE CLEANING - COMPLETELY REVISED
# ------------------------------------------------
# 1. First, look for any standard dividers that might separate thinking from lyrics
divider_patterns = [
r'Here are the lyrics:',
r'Here is my song:',
r'The lyrics:',
r'My lyrics:',
r'Song lyrics:',
r'\*\*\*+',
r'===+',
r'---+',
r'```',
r'Lyrics:'
]
for pattern in divider_patterns:
matches = re.finditer(pattern, lyrics, re.IGNORECASE)
for match in matches:
# Keep only content after the divider
lyrics = lyrics[match.end():].strip()
# 2. Remove thinking tags completely before splitting into lines
lyrics = re.sub(r'.*?', '', lyrics, flags=re.DOTALL)
lyrics = re.sub(r'\[thinking\].*?\[/thinking\]', '', lyrics, flags=re.DOTALL)
lyrics = re.sub(r'', '', lyrics, flags=re.DOTALL)
lyrics = re.sub(r'', '', lyrics, flags=re.DOTALL)
lyrics = re.sub(r'\[thinking\]', '', lyrics, flags=re.DOTALL)
lyrics = re.sub(r'\[/thinking\]', '', lyrics, flags=re.DOTALL)
# 3. Split text into lines for aggressive line-by-line filtering
lines = lyrics.strip().split('\n')
clean_lines = []
# 4. Define comprehensive patterns for non-lyrical content
non_lyric_patterns = [
# Meta-commentary
r'^(note|thinking|thoughts|let me|i will|i am going|i would|i can|i need to|i have to|i should|let\'s|here|now)',
r'^(first|second|third|next|finally|importantly|remember|so|ok|okay|as requested|as asked|considering)',
# Explanations
r'syllable[s]?|phrase|rhythm|beats?|tempo|bpm|instruction|follow|alignment|match|corresponding',
r'verses?|chorus|bridge|section|stanza|part|template|format|pattern|example',
r'requirements?|guidelines?|song structure|stressed|unstressed',
# Technical language
r'generated|output|result|provide|create|write|draft|version',
# Annotations and numbering
r'^line \d+|^\d+[\.\):]|^\[\w+\]|^[\*\-\+] ',
# Questions or analytical statements
r'\?$|analysis|evaluate|review|check|ensure',
# Instruction-like statements
r'make sure|please note|important|notice|pay attention'
]
# 5. Identify which lines are likely actual lyrics vs non-lyrics
for line in lines:
line = line.strip()
# Skip empty lines or lines with just spaces/tabs
if not line or line.isspace():
continue
# Skip lines that match any non-lyric pattern
should_skip = False
for pattern in non_lyric_patterns:
if re.search(pattern, line.lower()):
should_skip = True
break
if should_skip:
continue
# Skip section headers
if (line.startswith('[') and ']' in line) or (line.startswith('(') and ')' in line and len(line) < 20):
continue
# Skip lines that look like annotations (not prose-like)
if ':' in line and not any(word in line.lower() for word in ['like', 'when', 'where', 'how', 'why', 'what']):
if len(line.split(':')[0]) < 15: # Short prefixes followed by colon are likely annotations
continue
# Skip very short lines that aren't likely to be lyrics (unless it's just a few words which could be valid)
if len(line) < 3:
continue
# Skip lines that are numbered or bulleted
if re.match(r'^\d+\.|\(#\d+\)|\d+\)', line):
continue
# Skip markdown-style emphasis or headers
if re.match(r'^#{1,6} |^\*\*|^__', line):
continue
# Skip lines with think tags
if '' in line.lower() or '' in line.lower() or '[thinking]' in line.lower() or '[/thinking]' in line.lower():
continue
# Add this line as it passed all filters
clean_lines.append(line)
# 6. Additional block-level filters for common patterns
# Check beginning of lyrics for common prefixes
if clean_lines and any(clean_lines[0].lower().startswith(prefix) for prefix in
['here are', 'these are', 'below are', 'following are']):
clean_lines = clean_lines[1:] # Skip the first line
# 7. Process blocks of lines to detect explanation blocks
if len(clean_lines) > 3:
# Check for explanation blocks at the beginning
first_three = ' '.join(clean_lines[:3]).lower()
if any(term in first_three for term in ['i will', 'i have created', 'i\'ll provide', 'i\'ll write']):
# This looks like an explanation, skip the first few lines
start_idx = 0
for i, line in enumerate(clean_lines):
if i >= 3 and not any(term in line.lower() for term in ['i will', 'created', 'write', 'provide']):
start_idx = i
break
clean_lines = clean_lines[start_idx:]
# Check for explanation blocks at the end
last_three = ' '.join(clean_lines[-3:]).lower()
if any(term in last_three for term in ['hope this', 'these lyrics', 'as you can see', 'this song', 'i have']):
# This looks like an explanation at the end, truncate
end_idx = len(clean_lines)
for i in range(len(clean_lines) - 1, max(0, len(clean_lines) - 4), -1):
if i < len(clean_lines) and not any(term in clean_lines[i].lower() for term in
['hope', 'these lyrics', 'as you can see', 'this song']):
end_idx = i + 1
break
clean_lines = clean_lines[:end_idx]
# 8. Cleanup - Remove remaining annotations or thinking
for i in range(len(clean_lines)):
# Remove trailing thoughts/annotations
clean_lines[i] = re.sub(r'\s+//.*$', '', clean_lines[i])
clean_lines[i] = re.sub(r'\s+\(.*?\)$', '', clean_lines[i])
# Remove thinking tags completely
clean_lines[i] = re.sub(r'.*?', '', clean_lines[i], flags=re.DOTALL)
clean_lines[i] = re.sub(r'\[thinking\].*?\[/thinking\]', '', clean_lines[i], flags=re.DOTALL)
clean_lines[i] = re.sub(r'', '', clean_lines[i])
clean_lines[i] = re.sub(r'', '', clean_lines[i])
clean_lines[i] = re.sub(r'\[thinking\]', '', clean_lines[i])
clean_lines[i] = re.sub(r'\[/thinking\]', '', clean_lines[i])
# Remove syllable count annotations
clean_lines[i] = re.sub(r'\s*\(\d+\s*syllables?\)', '', clean_lines[i])
# 9. Filter out any remaining empty lines after tag removal
clean_lines = [line for line in clean_lines if line.strip() and not line.isspace()]
# 10. NEW: Apply strict syllable enforcement - split or truncate lines that are too long
# This is a critical step to ensure no line exceeds our max syllable count
if lyric_templates:
max_allowed_syllables = min(7, max([t.get('max_expected', 6) for t in lyric_templates]))
else:
max_allowed_syllables = 6
clean_lines = enforce_syllable_limits(clean_lines, max_allowed_syllables)
# 11. NEW: Check for template copying or clichéd phrases
cliched_patterns = [
r'moonlight (shimmers?|falls?|dances?)',
r'shadows? (dance|play|fall|stretch)',
r'time slips? away',
r'whispers? (fade|in the)',
r'silence speaks',
r'stars? shine',
r'hearts? beat',
r'footsteps (fade|echo)',
r'gentle wind',
r'(old|empty) (roads?|chair)',
r'night (holds?|falls?)',
r'memories fade',
r'dreams (linger|drift)'
]
cliche_count = 0
for line in clean_lines:
for pattern in cliched_patterns:
if re.search(pattern, line.lower()):
cliche_count += 1
break
# Calculate percentage of clichéd lines
if clean_lines:
cliche_percentage = (cliche_count / len(clean_lines)) * 100
else:
cliche_percentage = 0
# 12. If we have lyric templates, ensure we have the correct number of lines
if lyric_templates:
num_required = len(lyric_templates)
# If we have too many lines, keep just the best ones
if len(clean_lines) > num_required:
# Keep the first num_required lines
clean_lines = clean_lines[:num_required]
# If we don't have enough lines, generate placeholders that fit the syllable count
while len(clean_lines) < num_required:
i = len(clean_lines)
if i < len(lyric_templates):
template = lyric_templates[i]
target_syllables = min(max_allowed_syllables - 1, (template.get('min_expected', 2) + template.get('max_expected', 6)) // 2)
# Generate more creative, contextual placeholders with specificity
# Avoid clichés like "moonlight shimmers" or "time slips away"
specific_placeholders = {
# 2-3 syllables - specific, concrete phrases
2: [
"Phone rings twice",
"Dogs bark loud",
"Keys dropped here",
"Train rolls by",
"Birds take flight"
],
# 3-4 syllables - specific contexts
3: [
"Coffee gets cold",
"Fan blades spin",
"Pages turn slow",
"Neighbors talk",
"Radio hums soft"
],
# 4-5 syllables - specific details
4: [
"Fingers tap table",
"Taxi waits in rain",
"Laptop screen blinks",
"Ring left on sink",
"Church bells ring loud"
],
# 5-6 syllables - context rich
5: [
"Letters with no stamps",
"Watch shows wrong time",
"Jeans with torn knees",
"Dog barks next door",
"Smoke alarm beeps"
]
}
# Make theme and emotion specific placeholders to add to the list
theme_specific = []
if theme.lower() in ["love", "relationship", "romance"]:
theme_specific = ["Lipstick on glass", "Text left on read", "Scent on your coat"]
elif theme.lower() in ["loss", "grief", "sadness"]:
theme_specific = ["Chair sits empty", "Photos face down", "Clothes in closet"]
elif theme.lower() in ["hope", "inspiration", "triumph"]:
theme_specific = ["Seeds start to grow", "Finish line waits", "New day breaks through"]
# Get the closest matching syllable group
closest_group = min(specific_placeholders.keys(), key=lambda k: abs(k - target_syllables))
# Create pool of available placeholders from both specific and theme specific options
all_placeholders = specific_placeholders[closest_group] + theme_specific
# Choose a placeholder that hasn't been used yet
available_placeholders = [p for p in all_placeholders if p not in clean_lines]
if available_placeholders:
# Use modulo for more variation
idx = (i * 17 + len(clean_lines) * 13) % len(available_placeholders)
placeholder = available_placeholders[idx]
else:
# If we've used all placeholders, create something random and specific
subjects = ["Car", "Dog", "Kid", "Clock", "Phone", "Tree", "Book", "Door", "Light"]
verbs = ["waits", "moves", "stops", "falls", "breaks", "turns", "sleeps"]
# Ensure randomness with seed that changes with each call
import random
random.seed(len(clean_lines) * 27 + i * 31)
subj = random.choice(subjects)
verb = random.choice(verbs)
placeholder = f"{subj} {verb}"
else:
placeholder = "Page turns slow"
clean_lines.append(placeholder)
# Assemble final lyrics
final_lyrics = '\n'.join(clean_lines)
# Add a warning if we detected too many clichés
if cliche_percentage >= 40:
final_lyrics = f"""WARNING: These lyrics contain several overused phrases and clichés.
Try regenerating for more original content.
{final_lyrics}"""
# 13. Final sanity check - if we have nothing or garbage, return an error
if not final_lyrics or len(final_lyrics) < 10:
return "The model generated only thinking content but no actual lyrics. Please try again."
return final_lyrics
except Exception as e:
error_msg = f"Error generating lyrics: {str(e)}"
print(error_msg)
return error_msg
def analyze_lyrics_rhythm_match(lyrics, lyric_templates, genre="pop"):
"""Analyze how well the generated lyrics match the beat patterns and syllable requirements"""
if not lyric_templates or not lyrics:
return "No beat templates or lyrics available for analysis."
# Split lyrics into lines
lines = lyrics.strip().split('\n')
lines = [line for line in lines if line.strip()] # Remove empty lines
# Prepare analysis result
result = "### Beat & Syllable Match Analysis\n\n"
result += "| Line | Syllables | Target Range | Match | Stress Pattern |\n"
result += "| ---- | --------- | ------------ | ----- | -------------- |\n"
# Maximum number of lines to analyze (either all lines or all templates)
line_count = min(len(lines), len(lyric_templates))
# Track overall match statistics
total_matches = 0
total_range_matches = 0
total_stress_matches = 0
total_stress_percentage = 0
total_ideal_matches = 0
for i in range(line_count):
line = lines[i]
template = lyric_templates[i]
# Check match between line and template with genre awareness
check_result = beat_analyzer.check_syllable_stress_match(line, template, genre)
# Get match symbols
if check_result["close_to_ideal"]:
syllable_match = "✓" # Ideal or very close
elif check_result["within_range"]:
syllable_match = "✓*" # Within range but not ideal
else:
syllable_match = "✗" # Outside range
stress_match = "✓" if check_result["stress_matches"] else f"{int(check_result['stress_match_percentage']*100)}%"
# Update stats
if check_result["close_to_ideal"]:
total_matches += 1
total_ideal_matches += 1
elif check_result["within_range"]:
total_range_matches += 1
if check_result["stress_matches"]:
total_stress_matches += 1
total_stress_percentage += check_result["stress_match_percentage"]
# Create visual representation of the stress pattern
stress_visual = ""
for char in template['stress_pattern']:
if char == "S":
stress_visual += "X" # Strong
elif char == "M":
stress_visual += "x" # Medium
else:
stress_visual += "." # Weak
# Add line to results table
result += f"| {i+1} | {check_result['syllable_count']} | {check_result['min_expected']}-{check_result['max_expected']} | {syllable_match} | {stress_visual} |\n"
# Add summary statistics
if line_count > 0:
exact_match_rate = (total_matches / line_count) * 100
range_match_rate = ((total_matches + total_range_matches) / line_count) * 100
ideal_match_rate = (total_ideal_matches / line_count) * 100
stress_match_rate = (total_stress_matches / line_count) * 100
avg_stress_percentage = (total_stress_percentage / line_count) * 100
result += f"\n**Summary:**\n"
result += f"- Ideal or near-ideal syllable match rate: {exact_match_rate:.1f}%\n"
result += f"- Genre-appropriate syllable range match rate: {range_match_rate:.1f}%\n"
result += f"- Perfect stress pattern match rate: {stress_match_rate:.1f}%\n"
result += f"- Average stress pattern accuracy: {avg_stress_percentage:.1f}%\n"
result += f"- Overall rhythmic accuracy: {((range_match_rate + avg_stress_percentage) / 2):.1f}%\n"
# Analyze sentence flow across lines
sentence_flow_analysis = analyze_sentence_flow(lines)
result += f"\n**Sentence Flow Analysis:**\n"
result += f"- Connected thought groups: {sentence_flow_analysis['connected_groups']} detected\n"
result += f"- Average lines per thought: {sentence_flow_analysis['avg_lines_per_group']:.1f}\n"
result += f"- Flow quality: {sentence_flow_analysis['flow_quality']}\n"
# Add guidance on ideal distribution for syllables and sentence flow
result += f"\n**Syllable & Flow Guidance:**\n"
result += f"- Aim for {min([t.get('min_expected', 3) for t in lyric_templates])}-{max([t.get('max_expected', 7) for t in lyric_templates])} syllables per line\n"
result += f"- Break complete thoughts across 2-3 lines for natural flow\n"
result += f"- Connect your lyrics with sentence fragments that flow across lines\n"
result += f"- Use conjunctions, prepositions, and dependent clauses to connect lines\n"
# Add genre-specific notes
result += f"\n**Genre Notes ({genre}):**\n"
# Add appropriate genre notes based on genre
if genre.lower() == "pop":
result += "- Pop lyrics work well with thoughts spanning 2-3 musical phrases\n"
result += "- Create flow by connecting lines with transitions like 'as', 'when', 'through'\n"
elif genre.lower() == "rock":
result += "- Rock lyrics benefit from short phrases that build into complete thoughts\n"
result += "- Use line breaks strategically to emphasize key words\n"
elif genre.lower() == "country":
result += "- Country lyrics tell stories that flow naturally across multiple lines\n"
result += "- Connect narrative elements across phrases for authentic storytelling\n"
elif genre.lower() == "disco":
result += "- Disco lyrics work well with phrases that create rhythmic momentum\n"
result += "- Use line transitions that maintain energy and flow\n"
elif genre.lower() == "metal":
result += "- Metal lyrics can create intensity by breaking phrases at dramatic points\n"
result += "- Connect lines to build tension and release across measures\n"
else:
result += "- This genre works well with connected thoughts across multiple lines\n"
result += "- Aim for natural speech flow rather than complete thoughts per line\n"
return result
def analyze_sentence_flow(lines):
"""Analyze how well the lyrics create sentence flow across multiple lines"""
if not lines or len(lines) < 2:
return {
"connected_groups": 0,
"avg_lines_per_group": 0,
"flow_quality": "Insufficient lines to analyze"
}
# Simplified analysis looking for grammatical clues of sentence continuation
continuation_starters = [
'and', 'but', 'or', 'nor', 'for', 'yet', 'so', # Coordinating conjunctions
'as', 'when', 'while', 'before', 'after', 'since', 'until', 'because', 'although', 'though', # Subordinating conjunctions
'with', 'without', 'through', 'throughout', 'beyond', 'beneath', 'under', 'over', 'into', 'onto', # Prepositions
'to', 'from', 'by', 'at', 'in', 'on', 'of', # Common prepositions
'where', 'how', 'who', 'whom', 'whose', 'which', 'that', # Relative pronouns
'if', 'then', # Conditional connectors
]
# Check for lines that likely continue a thought from previous line
connected_lines = []
potential_groups = []
current_group = [0] # Start with first line
for i in range(1, len(lines)):
# Check if line starts with a continuation word
words = lines[i].lower().split()
# Empty line or no words
if not words:
if len(current_group) > 1: # Only consider groups of 2+ lines
potential_groups.append(current_group.copy())
current_group = [i]
continue
# Check first word for continuation clues
first_word = words[0].strip(',.!?;:')
if first_word in continuation_starters:
connected_lines.append(i)
current_group.append(i)
# Check for absence of capitalization as continuation clue
elif not first_word[0].isupper() and first_word[0].isalpha():
connected_lines.append(i)
current_group.append(i)
# Check if current line is very short (likely part of a continued thought)
elif len(words) <= 3 and i < len(lines) - 1:
# Look ahead to see if next line could be a continuation
if i+1 < len(lines):
next_words = lines[i+1].lower().split()
if next_words and next_words[0] in continuation_starters:
connected_lines.append(i)
current_group.append(i)
else:
# This might end a group
if len(current_group) > 1: # Only consider groups of 2+ lines
potential_groups.append(current_group.copy())
current_group = [i]
else:
# This likely starts a new thought
if len(current_group) > 1: # Only consider groups of 2+ lines
potential_groups.append(current_group.copy())
current_group = [i]
# Add the last group if it has multiple lines
if len(current_group) > 1:
potential_groups.append(current_group)
# Calculate metrics
connected_groups = len(potential_groups)
if connected_groups > 0:
avg_lines_per_group = sum(len(group) for group in potential_groups) / connected_groups
# Determine flow quality
if connected_groups >= len(lines) / 3 and avg_lines_per_group >= 2.5:
flow_quality = "Excellent - multiple connected thoughts across lines"
elif connected_groups >= len(lines) / 4 and avg_lines_per_group >= 2:
flow_quality = "Good - some connected thoughts across lines"
elif connected_groups > 0:
flow_quality = "Fair - limited connection between lines"
else:
flow_quality = "Poor - mostly independent lines"
else:
avg_lines_per_group = 0
flow_quality = "Poor - no connected thoughts detected"
return {
"connected_groups": connected_groups,
"avg_lines_per_group": avg_lines_per_group,
"flow_quality": flow_quality
}
def enforce_syllable_limits(lines, max_syllables=6):
"""
Enforce syllable limits by splitting or truncating lines that are too long.
Returns a modified list of lines where no line exceeds max_syllables.
"""
if not lines:
return []
result_lines = []
for line in lines:
words = line.split()
if not words:
continue
# Count syllables in the line
syllable_count = sum(beat_analyzer.count_syllables(word) for word in words)
# If within limits, keep the line as is
if syllable_count <= max_syllables:
result_lines.append(line)
continue
# Line is too long - we need to split or truncate it
current_line = []
current_syllables = 0
for word in words:
word_syllables = beat_analyzer.count_syllables(word)
# If adding this word would exceed the limit, start a new line
if current_syllables + word_syllables > max_syllables and current_line:
result_lines.append(" ".join(current_line))
current_line = [word]
current_syllables = word_syllables
else:
# Add the word to the current line
current_line.append(word)
current_syllables += word_syllables
# Don't forget the last line if there are words left
if current_line:
result_lines.append(" ".join(current_line))
return result_lines
# Create Gradio interface
def create_interface():
with gr.Blocks(title="Music Analysis & Lyrics Generator") as demo:
gr.Markdown("# Music Analysis & Lyrics Generator")
gr.Markdown("Upload a music file or record audio to analyze it and generate matching lyrics")
with gr.Row():
with gr.Column(scale=1):
audio_input = gr.Audio(
label="Upload or Record Audio",
type="filepath",
sources=["upload", "microphone"]
)
analyze_btn = gr.Button("Analyze and Generate Lyrics", variant="primary")
with gr.Column(scale=2):
with gr.Tab("Analysis"):
analysis_output = gr.Textbox(label="Music Analysis Results", lines=10)
with gr.Row():
tempo_output = gr.Number(label="Tempo (BPM)")
time_sig_output = gr.Textbox(label="Time Signature")
emotion_output = gr.Textbox(label="Primary Emotion")
theme_output = gr.Textbox(label="Primary Theme")
genre_output = gr.Textbox(label="Primary Genre")
with gr.Tab("Generated Lyrics"):
lyrics_output = gr.Textbox(label="Generated Lyrics", lines=20)
with gr.Tab("Beat Matching"):
beat_match_output = gr.Markdown(label="Beat & Syllable Matching Analysis")
# Set up event handlers
analyze_btn.click(
fn=process_audio,
inputs=[audio_input],
outputs=[analysis_output, lyrics_output, tempo_output, time_sig_output,
emotion_output, theme_output, genre_output, beat_match_output]
)
# Format supported genres for display
supported_genres_md = "\n".join([f"- {genre.capitalize()}" for genre in beat_analyzer.supported_genres])
gr.Markdown(f"""
## How it works
1. Upload or record a music file
2. The system analyzes tempo, beats, time signature and other musical features
3. It detects emotion, theme, and music genre
4. Using beat patterns and syllable stress analysis, it generates perfectly aligned lyrics
5. Each line of the lyrics is matched to the beat pattern of the corresponding musical phrase
## Supported Genres
**Note:** Lyrics generation is currently only supported for the following genres:
{supported_genres_md}
These genres have consistent syllable-to-beat patterns that work well with our algorithm.
For other genres, only music analysis will be provided.
""")
return demo
# Launch the app
demo = create_interface()
if __name__ == "__main__":
demo.launch()
else:
# For Hugging Face Spaces
app = demo