Loop-Architect / app.py
SaltProphet's picture
Update app.py
b1eee94 verified
import gradio as gr
import numpy as np
import librosa
import librosa.display
import soundfile as sf
import os
import tempfile
import zipfile
import time
import matplotlib
import matplotlib.pyplot as plt
from scipy import signal
from typing import Tuple, List, Any, Optional, Dict
import shutil
# Use a non-interactive backend for Matplotlib
matplotlib.use('Agg')
# --- CONSTANTS & DICTIONARIES ---
KEY_TO_CAMELOT = {
"C Maj": "8B", "G Maj": "9B", "D Maj": "10B", "A Maj": "11B", "E Maj": "12B",
"B Maj": "1B", "F# Maj": "2B", "Db Maj": "3B", "Ab Maj": "4B", "Eb Maj": "5B",
"Bb Maj": "6B", "F Maj": "7B",
"A Min": "8A", "E Min": "9A", "B Min": "10A", "F# Min": "11A", "C# Min": "12A",
"G# Min": "1A", "D# Min": "2A", "Bb Min": "3A", "F Min": "4A", "C Min": "5A",
"G Min": "6A", "D Min": "7A",
# Enharmonic equivalents
"Gb Maj": "2B", "Cb Maj": "7B", "A# Min": "3A", "D# Maj": "5B", "G# Maj": "4B"
}
# Fixed reverse mapping to avoid "lossy" inversion
CAMELOT_TO_KEY = {
"8B": "C Maj", "9B": "G Maj", "10B": "D Maj", "11B": "A Maj", "12B": "E Maj",
"1B": "B Maj", "2B": "F# Maj / Gb Maj", "3B": "Db Maj", "4B": "Ab Maj / G# Maj", "5B": "Eb Maj / D# Maj",
"6B": "Bb Maj", "7B": "F Maj / Cb Maj",
"8A": "A Min", "9A": "E Min", "10A": "B Min", "11A": "F# Min", "12A": "C# Min",
"1A": "G# Min", "2A": "D# Min", "3A": "Bb Min / A# Min", "4A": "F Min", "5A": "C Min",
"6A": "G Min", "7A": "D Min"
}
STEM_NAMES = ["vocals", "drums", "bass", "other", "guitar", "piano"]
# --- UTILITY FUNCTIONS ---
def freq_to_midi(freq: float) -> int:
"""Converts a frequency in Hz to a MIDI note number."""
if freq <= 0:
return 0
# C1 is ~32.7 Hz. Let's set a reasonable floor.
if freq < 32.0:
return 0
return int(round(69 + 12 * np.log2(freq / 440.0)))
def write_midi_file(notes_list: List[Tuple[int, float, float]], bpm: float, output_path: str):
"""
Writes a basic MIDI file from a list of notes.
Note: This is a simplified MIDI writer and may have issues.
Using a dedicated library like 'mido' is recommended for robust use.
"""
if not notes_list:
return
tempo_us_per_beat = int(60000000 / bpm)
division = 96 # Ticks per quarter note
seconds_per_tick = 60.0 / (bpm * division)
# Sort notes by start time
notes_list.sort(key=lambda x: x[1])
current_tick = 0
midi_events = []
# --- MIDI Track Header ---
# Set Tempo: FF 51 03 TTTTTT (TTTTTT = tempo_us_per_beat)
tempo_bytes = tempo_us_per_beat.to_bytes(3, 'big')
track_data = b'\x00\xFF\x51\x03' + tempo_bytes
# Set Time Signature: FF 58 04 NN DD CC BB (Using 4/4)
track_data += b'\x00\xFF\x58\x04\x04\x02\x18\x08'
# Set Track Name
track_data += b'\x00\xFF\x03\x0BLoopArchitect' # 11 chars
for note, start_sec, duration_sec in notes_list:
if note == 0:
continue
# Calculate delta time from last event
target_tick = int(round(start_sec / seconds_per_tick))
delta_tick = target_tick - current_tick
current_tick = target_tick
# Note On event (Channel 1, Velocity 100)
note_on = [0x90, note, 100]
track_data += encode_delta_time(delta_tick) + bytes(note_on)
# Note Off event (Channel 1, Velocity 0)
duration_ticks = int(round(duration_sec / seconds_per_tick))
if duration_ticks == 0:
duration_ticks = 1 # Minimum duration
note_off = [0x80, note, 0]
track_data += encode_delta_time(duration_ticks) + bytes(note_off)
current_tick += duration_ticks
# End of track
track_data += b'\x00\xFF\x2F\x00'
# --- MIDI File Header ---
# MThd, header_length (6), format (1), num_tracks (1), division
header = b'MThd' + (6).to_bytes(4, 'big') + (1).to_bytes(2, 'big') + (1).to_bytes(2, 'big') + division.to_bytes(2, 'big')
# MTrk, track_length, track_data
track_chunk = b'MTrk' + len(track_data).to_bytes(4, 'big') + track_data
midi_data = header + track_chunk
with open(output_path, 'wb') as f:
f.write(midi_data)
def encode_delta_time(ticks: int) -> bytes:
"""Encodes an integer tick value into MIDI variable-length quantity."""
buffer = ticks & 0x7F
ticks >>= 7
if ticks > 0:
buffer |= 0x80
while ticks > 0:
buffer = (buffer << 8) | ((ticks & 0x7F) | 0x80)
ticks >>= 7
buffer = (buffer & 0xFFFFFF7F) # Clear MSB of last byte
# Convert buffer to bytes
byte_list = []
while buffer > 0:
byte_list.insert(0, buffer & 0xFF)
buffer >>= 8
if not byte_list:
return b'\x00'
return bytes(byte_list)
else:
return bytes([buffer])
def get_harmonic_recommendations(key_str: str) -> str:
"""Calculates harmonically compatible keys based on the Camelot wheel."""
code = KEY_TO_CAMELOT.get(key_str, "N/A")
if code == "N/A":
return "N/A (Key not recognized or 'Unknown Key' detected.)"
try:
num = int(code[:-1])
mode = code[-1]
opposite_mode = 'B' if mode == 'A' else 'A'
num_plus_one = (num % 12) + 1
num_minus_one = 12 if num == 1 else num - 1
recs_codes = [
f"{num}{opposite_mode}", # e.g., 8A (A Min) -> 8B (C Maj)
f"{num_plus_one}{mode}", # e.g., 8A (A Min) -> 9A (E Min)
f"{num_minus_one}{mode}" # e.g., 8A (A Min) -> 7A (D Min)
]
rec_keys = [f"{CAMELOT_TO_KEY.get(r_code, f'Code {r_code}')} ({r_code})" for r_code in recs_codes]
return " | ".join(rec_keys)
except Exception as e:
print(f"Error calculating recommendations: {e}")
return "N/A (Error calculating recommendations.)"
def detect_key(y: np.ndarray, sr: int) -> str:
"""Analyzes the audio to determine the most likely musical key."""
try:
chroma = librosa.feature.chroma_stft(y=y, sr=sr)
chroma_sums = np.sum(chroma, axis=1)
# Avoid division by zero if audio is silent
if np.sum(chroma_sums) == 0:
return "Unknown Key"
chroma_norm = chroma_sums / np.sum(chroma_sums)
# Krumhansl-Schmuckler key-finding algorithm templates
major_template = np.array([6.35, 2.23, 3.48, 2.33, 4.38, 4.09, 2.52, 5.19, 2.39, 3.66, 2.29, 2.88])
minor_template = np.array([6.33, 2.68, 3.52, 5.38, 2.60, 3.53, 2.54, 4.75, 3.98, 2.69, 3.34, 3.17])
# Normalize templates
major_template /= np.sum(major_template)
minor_template /= np.sum(minor_template)
pitch_classes = ['C', 'C#', 'D', 'D#', 'E', 'F', 'F#', 'G', 'G#', 'A', 'A#', 'B']
major_correlations = [np.dot(chroma_norm, np.roll(major_template, i)) for i in range(12)]
best_major_index = np.argmax(major_correlations)
minor_correlations = [np.dot(chroma_norm, np.roll(minor_template, i)) for i in range(12)]
best_minor_index = np.argmax(minor_correlations)
if major_correlations[best_major_index] > minor_correlations[best_minor_index]:
return pitch_classes[best_major_index] + " Maj"
else:
return pitch_classes[best_minor_index] + " Min"
except Exception as e:
print(f"Key detection failed: {e}")
return "Unknown Key"
def apply_modulation(y: np.ndarray, sr: int, bpm: float, rate: str, pan_depth: float, level_depth: float) -> np.ndarray:
"""Applies tempo-synced LFOs for panning and volume modulation."""
if y.ndim == 0:
return y
if y.ndim == 1:
y = np.stack((y, y), axis=-1) # Convert to stereo
N = len(y)
duration_sec = N / sr
rate_map = {'1/2': 0.5, '1/4': 1, '1/8': 2, '1/16': 4}
beats_per_measure = rate_map.get(rate, 1)
# LFO frequency = (BPM / 60) * (beats_per_measure / 4.0) -- seems off.
# Let's redefine: LFO freq in Hz = (BPM / 60) * (1 / (4 / beats_per_measure))
# e.g., 1/4 rate at 120BPM = 2Hz. (120/60) * (1 / (4/1)) = 2 * (1/4) = 0.5Hz? No.
# 120 BPM = 2 beats/sec. 1/4 note = 1 beat. So LFO should be 2 Hz.
# 1/8 note = 4 Hz.
# 1/16 note = 8 Hz.
# 1/2 note = 1 Hz.
# Formula: (BPM / 60) * (rate_map_value / 4)
# 1/4 note: (120/60) * (1/4) = 0.5 Hz. Still wrong.
# Let's try: (BPM / 60) * (rate_map_value)
# 1/4 note @ 120BPM: (120/60) * 1 = 2 Hz. Correct.
# 1/8 note @ 120BPM: (120/60) * 2 = 4 Hz. Correct.
# 1/2 note @ 120BPM: (120/60) * 0.5 = 1 Hz. Correct.
lfo_freq_hz = (bpm / 60.0) * rate_map.get(rate, 1)
t = np.linspace(0, duration_sec, N, endpoint=False)
# Panning LFO (Sine wave, -1 to 1)
if pan_depth > 0:
pan_lfo = np.sin(2 * np.pi * lfo_freq_hz * t) * pan_depth
# L_mod/R_mod should be 0-1. (1-pan_lfo)/2 and (1+pan_lfo)/2 gives 0-1 range.
L_mod = (1 - pan_lfo) / 2.0
R_mod = (1 + pan_lfo) / 2.0
# This is amplitude panning, not constant power. Good enough.
y[:, 0] *= L_mod
y[:, 1] *= R_mod
# Level LFO (Tremolo) (Sine wave, 0 to 1)
if level_depth > 0:
level_lfo = (np.sin(2 * np.pi * lfo_freq_hz * t) + 1) / 2.0
# gain_multiplier ranges from (1-level_depth) to 1
gain_multiplier = (1 - level_depth) + (level_depth * level_lfo)
y[:, 0] *= gain_multiplier
y[:, 1] *= gain_multiplier
return y
def apply_normalization_dbfs(y: np.ndarray, target_dbfs: float) -> np.ndarray:
"""Applies peak normalization to match a target dBFS value."""
if target_dbfs >= 0:
return y # Don't normalize to 0dBFS or higher
current_peak_amp = np.max(np.abs(y))
if current_peak_amp < 1e-9: # Avoid division by zero on silence
return y
target_peak_amp = 10**(target_dbfs / 20.0)
gain = target_peak_amp / current_peak_amp
y_normalized = y * gain
# Clip just in case of floating point inaccuracies
y_normalized = np.clip(y_normalized, -1.0, 1.0)
return y_normalized
def apply_filter_modulation(y: np.ndarray, sr: int, bpm: float, rate: str, filter_type: str, freq: float, depth: float) -> np.ndarray:
"""Applies a tempo-synced LFO to a 2nd order Butterworth filter cutoff frequency."""
if depth == 0 or filter_type == "None":
return y
# Ensure stereo for LFO application
if y.ndim == 1:
y = np.stack((y, y), axis=-1)
if y.ndim == 0:
return y
N = len(y)
duration_sec = N / sr
# LFO Rate Calculation
rate_map = {'1/2': 0.5, '1/4': 1, '1/8': 2, '1/16': 4}
lfo_freq_hz = (bpm / 60.0) * rate_map.get(rate, 1)
t = np.linspace(0, duration_sec, N, endpoint=False)
# LFO: ranges from 0 to 1
lfo_value = (np.sin(2 * np.pi * lfo_freq_hz * t) + 1) / 2.0
# Modulate Cutoff Frequency: Cutoff = BaseFreq + (LFO * Depth)
cutoff_modulation = freq + (lfo_value * depth)
# Safety clip to prevent instability
nyquist = sr / 2.0
cutoff_modulation = np.clip(cutoff_modulation, 20.0, nyquist - 100.0) # Keep away from Nyquist
y_out = np.zeros_like(y)
# --- BUG FIX ---
# Was: filter_type.lower().replace('-pass', '') -> 'low' (ValueError)
# Now: filter_type.lower().replace('-pass', 'pass') -> 'lowpass' (Correct)
filter_type_b = filter_type.lower().replace('-pass', 'pass')
frame_size = 512 # Frame-based update for filter coefficients
if N < frame_size:
frame_size = N # Handle very short audio
# Apply filter channel by channel
for channel in range(y.shape[1]):
zi = signal.lfilter_zi(*signal.butter(2, 20.0, btype=filter_type_b, fs=sr))
for frame_start in range(0, N, frame_size):
frame_end = min(frame_start + frame_size, N)
if frame_start == frame_end: continue # Skip empty frames
frame = y[frame_start:frame_end, channel]
# Use the average LFO cutoff for the frame
avg_cutoff = np.mean(cutoff_modulation[frame_start:frame_end])
# Calculate 2nd order Butterworth filter coefficients
try:
b, a = signal.butter(2, avg_cutoff, btype=filter_type_b, fs=sr)
except ValueError as e:
print(f"Butterworth filter error: {e}. Using last good coefficients.")
# This can happen if avg_cutoff is bad, though we clip it.
# If it still fails, we just re-use the last good b, a.
# In the first frame, this is not robust.
if 'b' not in locals():
b, a = signal.butter(2, 20.0, btype=filter_type_b, fs=sr) # Failsafe
# Apply filter to the frame, updating the state `zi`
filtered_frame, zi = signal.lfilter(b, a, frame, zi=zi)
y_out[frame_start:frame_end, channel] = filtered_frame
return y_out
def apply_crossfade(y: np.ndarray, fade_samples: int) -> np.ndarray:
"""Applies a linear fade-in and fade-out to a clip."""
if fade_samples == 0:
return y
N = len(y)
fade_samples = min(fade_samples, N // 2) # Fade can't be longer than half the clip
if fade_samples == 0:
return y # Clip is too short to fade
# Create fade ramps
fade_in = np.linspace(0, 1, fade_samples)
fade_out = np.linspace(1, 0, fade_samples)
y_out = y.copy()
# Apply fades (handling mono/stereo)
if y.ndim == 1:
y_out[:fade_samples] *= fade_in
y_out[-fade_samples:] *= fade_out
else:
y_out[:fade_samples, :] *= fade_in[:, np.newaxis]
y_out[-fade_samples:, :] *= fade_out[:, np.newaxis]
return y_out
def apply_envelope(y: np.ndarray, sr: int, attack_gain_db: float, sustain_gain_db: float) -> np.ndarray:
"""Applies a simple attack/sustain gain envelope to one-shots."""
N = len(y)
if N == 0:
return y
# Simple fixed attack time of 10ms
attack_time_sec = 0.01
attack_samples = min(int(attack_time_sec * sr), N // 2)
start_gain = 10**(attack_gain_db / 20.0)
end_gain = 10**(sustain_gain_db / 20.0)
# Envelope: Linear ramp from start_gain to end_gain over attack_samples, then hold end_gain
envelope = np.ones(N) * end_gain
if attack_samples > 0:
attack_ramp = np.linspace(start_gain, end_gain, attack_samples)
envelope[:attack_samples] = attack_ramp
# Apply envelope (handling mono/stereo)
if y.ndim == 1:
y_out = y * envelope
else:
y_out = y * envelope[:, np.newaxis]
return y_out
# --- CORE PROCESSING FUNCTIONS ---
def separate_stems(audio_file_path: str) -> Tuple[
Optional[Tuple[int, np.ndarray]],
Optional[Tuple[int, np.ndarray]],
Optional[Tuple[int, np.ndarray]],
Optional[Tuple[int, np.ndarray]],
Optional[Tuple[int, np.ndarray]],
Optional[Tuple[int, np.ndarray]],
float, str, str
]:
"""
Simulates stem separation and detects BPM and Key.
Returns Gradio Audio tuples (sr, data) for each stem.
"""
if audio_file_path is None:
raise gr.Error("No audio file uploaded!")
try:
# Load audio
y_orig, sr_orig = librosa.load(audio_file_path, sr=None, mono=False)
# Ensure stereo for processing
if y_orig.ndim == 1:
y_orig = np.stack([y_orig, y_orig], axis=-1)
# librosa.load with mono=False may return (N,) for mono files,
# or (2, N). Need to ensure (N, 2) or (N,)
if y_orig.ndim == 2 and y_orig.shape[0] < y_orig.shape[1]:
y_orig = y_orig.T # Transpose to (N, 2)
y_mono = librosa.to_mono(y_orig)
# Detect tempo and key
tempo, _ = librosa.beat.beat_track(y=y_mono, sr=sr_orig)
detected_bpm = 120.0 if tempo is None or tempo.size == 0 or tempo[0] == 0 else float(np.round(tempo[0]))
detected_key = detect_key(y_mono, sr_orig)
harmonic_recs = get_harmonic_recommendations(detected_key)
# Create mock separated stems
# In a real app, you'd use Demucs, Spleeter, etc.
# Here, we just return the original audio for each stem for demo purposes.
stems_data: Dict[str, Optional[Tuple[int, np.ndarray]]] = {}
# Convert to int16 for Gradio Audio component
y_int16 = (y_orig * 32767).astype(np.int16)
for name in STEM_NAMES:
# We give each stem the full audio for this demo
stems_data[name] = (sr_orig, y_int16.copy())
return (
stems_data["vocals"], stems_data["drums"], stems_data["bass"], stems_data["other"],
stems_data["guitar"], stems_data["piano"],
detected_bpm, detected_key, harmonic_recs
)
except Exception as e:
print(f"Error processing audio: {e}")
import traceback
traceback.print_exc()
raise gr.Error(f"Error processing audio: {str(e)}")
def generate_waveform_preview(y: np.ndarray, sr: int, stem_name: str, temp_dir: str) -> str:
"""Generates a Matplotlib image showing the waveform."""
img_path = os.path.join(temp_dir, f"{stem_name}_preview.png")
plt.figure(figsize=(10, 3))
y_display = librosa.to_mono(y.T) if y.ndim > 1 and y.shape[0] < y.shape[1] else y
y_display = librosa.to_mono(y) if y.ndim > 1 else y
librosa.display.waveshow(y_display, sr=sr, x_axis='time', color="#4a7098")
plt.title(f"{stem_name} Waveform (Processed)")
plt.ylabel("Amplitude")
plt.tight_layout()
plt.savefig(img_path)
plt.close()
return img_path
def slice_stem_real(
stem_audio_tuple: Optional[Tuple[int, np.ndarray]],
loop_choice: str,
sensitivity: float,
stem_name: str,
manual_bpm: float,
time_signature: str,
crossfade_ms: int,
transpose_semitones: int,
detected_key: str,
pan_depth: float,
level_depth: float,
modulation_rate: str,
target_dbfs: float,
attack_gain: float,
sustain_gain: float,
filter_type: str,
filter_freq: float,
filter_depth: float
) -> Tuple[List[str], Optional[str]]:
"""
Slices a single stem and applies transformations.
Returns a list of filepaths and a path to a preview image.
"""
if stem_audio_tuple is None:
return [], None
try:
sample_rate, y_int = stem_audio_tuple
# Convert from int16 array back to float
y = y_int.astype(np.float32) / 32767.0
if y.ndim == 0 or len(y) == 0:
return [], None
# --- 1. PITCH SHIFTING (if enabled) ---
if transpose_semitones != 0:
y = librosa.effects.pitch_shift(y, sr=sample_rate, n_steps=transpose_semitones)
# --- 2. FILTER MODULATION ---
if filter_depth > 0 and filter_type != "None":
y = apply_filter_modulation(y, sample_rate, manual_bpm, modulation_rate, filter_type, filter_freq, filter_depth)
# --- 3. PAN/LEVEL MODULATION ---
normalized_pan_depth = pan_depth / 100.0
normalized_level_depth = level_depth / 100.0
if normalized_pan_depth > 0 or normalized_level_depth > 0:
y = apply_modulation(y, sample_rate, manual_bpm, modulation_rate, normalized_pan_depth, normalized_level_depth)
# --- 4. NORMALIZATION ---
if target_dbfs < 0:
y = apply_normalization_dbfs(y, target_dbfs)
# --- 5. DETERMINE BPM & KEY ---
bpm_int = int(round(manual_bpm))
key_tag = "UnknownKey"
if detected_key != "Unknown Key":
key_tag = detected_key.replace(" ", "")
if transpose_semitones != 0:
root, mode = detected_key.split(" ")
pitch_classes = ['C', 'C#', 'D', 'D#', 'E', 'F', 'F#', 'G', 'G#', 'A', 'A#', 'B']
try:
current_index = pitch_classes.index(root)
new_index = (current_index + transpose_semitones) % 12
new_key_root = pitch_classes[new_index]
key_tag = f"{new_key_root}{mode}Shift"
except ValueError:
key_tag = f"Shifted{transpose_semitones}" # Fallback
# --- 6. MIDI GENERATION (Melodic Stems) ---
output_files = []
loops_dir = tempfile.mkdtemp()
is_melodic = stem_name in ["vocals", "bass", "guitar", "piano", "other"]
if is_melodic and ("Bar Loops" in loop_choice):
try:
y_mono_for_midi = librosa.to_mono(y)
# Use piptrack for pitch detection
pitches, magnitudes = librosa.piptrack(y=y_mono_for_midi, sr=sample_rate)
# Get the dominant pitch at each frame
main_pitch_line = np.zeros(pitches.shape[1])
for t in range(pitches.shape[1]):
index = magnitudes[:, t].argmax()
main_pitch_line[t] = pitches[index, t]
notes_list = []
i = 0
hop_length = 512 # Default hop for piptrack
while i < len(main_pitch_line):
current_freq = main_pitch_line[i]
current_midi = freq_to_midi(current_freq)
if current_midi == 0: # Skip silence/unpitched
i += 1
continue
# Find end of this note
j = i
while j < len(main_pitch_line) and freq_to_midi(main_pitch_line[j]) == current_midi:
j += 1
duration_frames = j - i
# Only add notes that are long enough (e.g., > 2 frames)
if duration_frames >= 2:
start_sec = librosa.frames_to_time(i, sr=sample_rate, hop_length=hop_length)
duration_sec = librosa.frames_to_time(duration_frames, sr=sample_rate, hop_length=hop_length)
notes_list.append((current_midi, start_sec, duration_sec))
i = j
if notes_list:
full_stem_midi_path = os.path.join(loops_dir, f"{stem_name}_MELODY_{key_tag}_{bpm_int}BPM.mid")
write_midi_file(notes_list, manual_bpm, full_stem_midi_path)
output_files.append(full_stem_midi_path)
except Exception as e:
print(f"MIDI generation failed for {stem_name}: {e}")
# --- 7. CALCULATE TIMING & SLICING ---
beats_per_bar = 4
if time_signature == "3/4":
beats_per_bar = 3
if "Bar Loops" in loop_choice:
bars = int(loop_choice.split(" ")[0])
loop_type_tag = f"{bars}Bar"
loop_duration_samples = int((60.0 / manual_bpm * beats_per_bar * bars) * sample_rate)
fade_samples = int((crossfade_ms / 1000.0) * sample_rate)
if loop_duration_samples > 0 and len(y) > loop_duration_samples:
num_loops = len(y) // loop_duration_samples
for i in range(min(num_loops, 16)): # Limit to 16 loops
start_sample = i * loop_duration_samples
end_sample = min(start_sample + loop_duration_samples, len(y))
slice_data = y[start_sample:end_sample]
# Apply crossfade
slice_data = apply_crossfade(slice_data, fade_samples)
filename = os.path.join(loops_dir, f"{stem_name}_{loop_type_tag}_{i+1:03d}_{key_tag}_{bpm_int}BPM.wav")
sf.write(filename, slice_data, sample_rate, subtype='PCM_16')
output_files.append(filename)
elif "One-Shots" in loop_choice:
loop_type_tag = "OneShot"
y_mono_for_onsets = librosa.to_mono(y)
# IMPLEMENTED: Use sensitivity to find onsets
# Adjust 'wait' and 'delta' based on sensitivity (0-1)
# Higher sensitivity = lower delta, shorter wait
delta = 0.5 * (1.0 - sensitivity) # 0.0 -> 0.5
wait_sec = 0.1 * (1.0 - sensitivity) # 0.0 -> 0.1
wait_samples = int(wait_sec * sample_rate / 512) # in frames
onset_frames = librosa.onset.onset_detect(
y=y_mono_for_onsets,
sr=sample_rate,
units='frames',
backtrack=True,
delta=delta,
wait=wait_samples
)
onset_samples = librosa.frames_to_samples(onset_frames)
# Add end of file as the last "onset"
onset_samples = np.append(onset_samples, len(y))
for i in range(min(len(onset_samples) - 1, 40)): # Limit to 40 slices
start_sample = onset_samples[i]
end_sample = onset_samples[i+1]
slice_data = y[start_sample:end_sample]
if len(slice_data) < 100: # Skip tiny fragments
continue
# IMPLEMENTED: Apply attack/sustain envelope
slice_data = apply_envelope(slice_data, sample_rate, attack_gain, sustain_gain)
# Apply short fade-out to prevent clicks
slice_data = apply_crossfade(slice_data, int(0.005 * sample_rate)) # 5ms fade
filename = os.path.join(loops_dir, f"{stem_name}_{loop_type_tag}_{i+1:03d}_{key_tag}_{bpm_int}BPM.wav")
sf.write(filename, slice_data, sample_rate, subtype='PCM_16')
output_files.append(filename)
# --- 8. VISUALIZATION GENERATION ---
img_path = generate_waveform_preview(y, sample_rate, stem_name, loops_dir)
# Clean up the temp dir for the *next* run
# Gradio File components need the files to exist, so we don't delete loops_dir yet
# A more robust solution would use gr.TempFile() or manage cleanup
return output_files, img_path
except Exception as e:
print(f"Error processing stem {stem_name}: {e}")
import traceback
traceback.print_exc()
return [], None # Return empty on error
def slice_all_and_zip(
vocals_audio: Optional[Tuple[int, np.ndarray]],
drums_audio: Optional[Tuple[int, np.ndarray]],
bass_audio: Optional[Tuple[int, np.ndarray]],
other_audio: Optional[Tuple[int, np.ndarray]],
guitar_audio: Optional[Tuple[int, np.ndarray]],
piano_audio: Optional[Tuple[int, np.ndarray]],
loop_choice: str,
sensitivity: float,
manual_bpm: float,
time_signature: str,
crossfade_ms: int,
transpose_semitones: int,
detected_key: str,
pan_depth: float,
level_depth: float,
modulation_rate: str,
target_dbfs: float,
attack_gain: float,
sustain_gain: float,
filter_type: str,
filter_freq: float,
filter_depth: float,
progress: gr.Progress
) -> Optional[str]:
"""Slices all available stems and packages them into a ZIP file."""
try:
stems_to_process = {
"vocals": vocals_audio, "drums": drums_audio, "bass": bass_audio,
"other": other_audio, "guitar": guitar_audio, "piano": piano_audio
}
# Filter out None stems
valid_stems = {name: data for name, data in stems_to_process.items() if data is not None}
if not valid_stems:
raise gr.Error("No stems to process! Please separate stems first.")
# Create temporary directory for all outputs
temp_dir = tempfile.mkdtemp()
zip_path = os.path.join(temp_dir, "Loop_Architect_Pack.zip")
all_sliced_files = []
# Use progress tracker
progress(0, desc="Starting...")
num_stems = len(valid_stems)
for i, (name, data) in enumerate(valid_stems.items()):
progress((i+1)/num_stems, desc=f"Slicing {name}...")
# Process stem
sliced_files, _ = slice_stem_real(
data, loop_choice, sensitivity, name,
manual_bpm, time_signature, crossfade_ms, transpose_semitones, detected_key,
pan_depth, level_depth, modulation_rate, target_dbfs,
attack_gain, sustain_gain, filter_type, filter_freq, filter_depth
)
all_sliced_files.extend(sliced_files)
progress(0.9, desc="Zipping files...")
with zipfile.ZipFile(zip_path, 'w', zipfile.ZIP_DEFLATED) as zf:
for file_path in all_sliced_files:
if not file_path: continue
# Create a sane folder structure in the ZIP
file_type = os.path.splitext(file_path)[1][1:].upper() # WAV or MID
arcname = os.path.join(file_type, os.path.basename(file_path))
zf.write(file_path, arcname)
progress(1.0, desc="Done!")
# Clean up individual slice files (but not the zip dir)
for file_path in all_sliced_files:
if file_path and os.path.exists(file_path):
os.remove(file_path)
return zip_path
except Exception as e:
print(f"Error creating ZIP: {e}")
import traceback
traceback.print_exc()
raise gr.Error(f"Error creating ZIP: {str(e)}")
# --- GRADIO INTERFACE ---
with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", secondary_hue="red")) as demo:
gr.Markdown("# 🎵 Loop Architect (Pro Edition)")
gr.Markdown("Upload any song to separate it into stems, detect musical attributes, and then slice and tag the stems for instant use in a DAW.")
# State variables
detected_bpm_state = gr.State(value=120.0)
detected_key_state = gr.State(value="Unknown Key")
harmonic_recs_state = gr.State(value="---")
# Outputs for each stem (as gr.Audio tuples)
vocals_audio = gr.Audio(visible=False, type="numpy")
drums_audio = gr.Audio(visible=False, type="numpy")
bass_audio = gr.Audio(visible=False, type="numpy")
other_audio = gr.Audio(visible=False, type="numpy")
guitar_audio = gr.Audio(visible=False, type="numpy")
piano_audio = gr.Audio(visible=False, type="numpy")
stem_audio_outputs = [vocals_audio, drums_audio, bass_audio, other_audio, guitar_audio, piano_audio]
with gr.Row():
with gr.Column(scale=1):
# --- INPUT COLUMN ---
gr.Markdown("### 1. Upload & Analyze")
audio_input = gr.Audio(label="Upload Song", type="filepath")
separate_button = gr.Button("Separate Stems & Analyze", variant="primary")
with gr.Accordion("Global Musical Settings", open=True):
manual_bpm_input = gr.Number(label="BPM", value=120.0, step=0.1, interactive=True)
key_display = gr.Textbox(label="Detected Key", value="Unknown Key", interactive=False)
harmonic_recs_display = gr.Textbox(label="Harmonic Recommendations", value="---", interactive=False)
transpose_semitones = gr.Slider(label="Transpose (Semitones)", minimum=-12, maximum=12, value=0, step=1)
time_signature = gr.Radio(label="Time Signature", choices=["4/4", "3/4"], value="4/4")
with gr.Accordion("Global Slicing Settings", open=True):
loop_choice = gr.Radio(label="Loop Type", choices=["1 Bar Loops", "2 Bar Loops", "4 Bar Loops", "One-Shots"], value="4 Bar Loops")
sensitivity = gr.Slider(label="One-Shot Sensitivity", minimum=0.0, maximum=1.0, value=0.5, info="Higher = more slices")
crossfade_ms = gr.Slider(label="Loop Crossfade (ms)", minimum=0, maximum=50, value=10, step=1)
with gr.Accordion("Global FX Settings", open=False):
target_dbfs = gr.Slider(label="Normalize Peak to (dBFS)", minimum=-24.0, maximum=-0.0, value=-1.0, step=0.1, info="-0.0 = Off")
gr.Markdown("---")
gr.Markdown("**LFO Modulation (Pan/Level)**")
modulation_rate = gr.Radio(label="Modulation Rate", choices=["1/2", "1/4", "1/8", "1/16"], value="1/4")
pan_depth = gr.Slider(label="Pan Depth (%)", minimum=0, maximum=100, value=0, step=1)
level_depth = gr.Slider(label="Level Depth (%)", minimum=0, maximum=100, value=0, step=1, info="Tremolo effect")
gr.Markdown("---")
gr.Markdown("**LFO Modulation (Filter)**")
filter_type = gr.Radio(label="Filter Type", choices=["None", "Low-pass", "High-pass"], value="None")
filter_freq = gr.Slider(label="Filter Base Freq (Hz)", minimum=20, maximum=10000, value=5000, step=100)
filter_depth = gr.Slider(label="Filter Mod Depth (Hz)", minimum=0, maximum=10000, value=0, step=100, info="LFO amount")
gr.Markdown("---")
gr.Markdown("**One-Shot Shaping**")
attack_gain = gr.Slider(label="Attack Gain (dB)", minimum=-24.0, maximum=6.0, value=0.0, step=0.5, info="Gain at start of transient")
sustain_gain = gr.Slider(label="Sustain Gain (dB)", minimum=-24.0, maximum=6.0, value=0.0, step=0.5, info="Gain for note body")
gr.Markdown("### 3. Generate Pack")
slice_all_button = gr.Button("SLICE ALL & GENERATE PACK", variant="primary")
zip_file_output = gr.File(label="Download Your Loop Pack")
with gr.Column(scale=2):
# --- OUTPUT COLUMN ---
gr.Markdown("### 2. Review Stems & Slices")
with gr.Tabs():
# Create a tab for each stem
for i, name in enumerate(STEM_NAMES):
with gr.Tab(name.capitalize()):
with gr.Row():
# The (hidden) audio output for this stem
stem_audio_component = stem_audio_outputs[i]
# Visible components
preview_image = gr.Image(label="Processed Waveform", interactive=False)
slice_files = gr.Files(label="Generated Slices & MIDI", interactive=False)
# Add a button to slice just this one stem
slice_one_button = gr.Button(f"Slice This {name.capitalize()} Stem")
# Gather all global settings as inputs
all_settings = [
loop_choice, sensitivity, manual_bpm_input, time_signature, crossfade_ms,
transpose_semitones, detected_key_state, pan_depth, level_depth, modulation_rate,
target_dbfs, attack_gain, sustain_gain, filter_type, filter_freq, filter_depth
]
# Wire up the "Slice One" button
slice_one_button.click(
fn=slice_stem_real,
inputs=[stem_audio_component, gr.State(value=name)] + all_settings,
outputs=[slice_files, preview_image]
)
# --- EVENT LISTENERS ---
# 1. "Separate Stems" button click
separate_button.click(
fn=separate_stems,
inputs=[audio_input],
outputs=stem_audio_outputs + [detected_bpm_state, detected_key_state, harmonic_recs_state]
)
# 2. When BPM state changes, update the visible input box
detected_bpm_state.change(
fn=lambda x: x,
inputs=[detected_bpm_state],
outputs=[manual_bpm_input]
)
# 3. When Key state changes, update the visible text boxes
detected_key_state.change(
fn=lambda x: x,
inputs=[detected_key_state],
outputs=[key_display]
)
harmonic_recs_state.change(
fn=lambda x: x,
inputs=[harmonic_recs_state],
outputs=[harmonic_recs_display]
)
# 4. "SLICE ALL" button click
slice_all_button.click(
fn=slice_all_and_zip,
inputs=stem_audio_outputs + all_settings,
outputs=[zip_file_output]
)
if __name__ == "__main__":
demo.launch(debug=True)