PRANJAL KAR
FEAT: Added Tracing for user selections
c893884
import os
import shutil
import json
import tempfile
import gradio as gr
from pydub import AudioSegment, silence
from pydub.effects import low_pass_filter, high_pass_filter
from tqdm import tqdm
from groq import Groq
import logging
from langsmith import traceable
from langchain.callbacks import LangChainTracer
from dotenv import load_dotenv
import random
from temp_choose import *
load_dotenv()
logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
logger.addHandler(logging.StreamHandler())
logger.addHandler(logging.FileHandler("log.txt"))
@traceable(run_type="llm", name="groq_call")
def make_groq_call(stems, song_name, p, section_type=None, bpm=120, bars=16):
"""
Make a call to the Groq API to get music production instructions.
Args:
stems (list): List of available stem files
song_name (str): Name of the song
p (float): Variation parameter (0-1)
section_type (str, optional): Specific section to generate variants for
bpm (int): Beats per minute
bars (int): Number of bars
Returns:
dict: JSON response with production instructions
"""
client = Groq(api_key=os.getenv("GROQ_API_KEY"))
# Customize prompt based on whether we're generating a full track or section variants
if section_type:
system_content = """You are a very experienced music producer and analyst, with a deep understanding of music theory and production techniques and arrangement, with specific expertise in creating popular EDM/ dance music.
You are given a set of audio stems and asked to create multiple variants of a specific section of a track.
Your task is to provide detailed instructions on how to arrange and process the stems for each variant.
You will be given audio stems and asked to create multiple variants of a specific section of a track.
For each variant, return detailed instructions on how to arrange and process the stems.
Be creative and make each variant sound distinct while maintaining a coherent musical style."""
user_content = f"""I need 4 different variants for the {section_type} section of a track named "{song_name}".
Available stems: {stems}
BPM: {bpm}
Length of the section: {bars} bars
For each variant, please provide specific instructions on:
1. Which stems to include
2. What audio operations to apply (filters, fades, etc.)
3. How the stems should be arranged
Make the variants diverse but coherent, with variation level p={p} (0=minimal variation, 1=maximum variation).
Return your response as a JSON object with this structure:
{{
"variant1": {{
"stems": ["stem1.wav", "stem2.wav"],
"operations": [
{{"stem": "stem1.wav", "operation": "low_pass_filter", "value": 500}},
{{"stem": "stem2.wav", "operation": "fade_in", "value": 1000}}
],
"overlay": true,
"description": "A brief description of this variant"
}},
"variant2": {{ ... }},
"variant3": {{ ... }},
"variant4": {{ ... }}
}}
"""
else:
system_content = """You are a very experienced music producer and analyst. For a given audio folder, that has the instruments, are combined to make a loop, and make some variations of it as well. After analyzing the code, you are supposed to return the code containing the different functions for producing the full track."""
user_content = f"""Now, you have a new song {song_name}, which has the following contents:
{stems}
BPM: {bpm}
Bars: {bars}
Return the code as per discussed in example. Make proper arrangements, for best groovy music. Create 3 variations and return code in JSON. And make sure to use as many instruments possible in each variation.
NOTE: And at least once, the MAIN loop should have ALL stems.
Like the variations should be like the main full loop, with some adjustments according to value of p={p} (p will remain in between 0-1; 0 means no variation in loop and 1 means high variation in loop), and then another variation can be a two times repeat of the full loop, with some effects in second time. Make intro better, with some more instruments.
The main loop should have ALL wav files, don't exclude any please. Return proper JSON, with all the keys and values.
"""
completion = client.chat.completions.create(
model="gemma2-9b-it",
messages=[
{"role": "system", "content": system_content},
{"role": "user", "content": user_content},
],
temperature=1,
top_p=1,
stream=False,
response_format={"type": "json_object"},
stop=None,
)
print(completion.choices[0].message.content)
return json.loads(completion.choices[0].message.content)
def rename_files_remove_spaces(folder):
"""Rename all files in the folder by removing spaces from filenames"""
files_renamed = 0
for file in os.listdir(folder):
if " " in file:
old_path = os.path.join(folder, file)
new_file = file.replace(" ", "")
new_path = os.path.join(folder, new_file)
# Only rename if the new file doesn't already exist
if not os.path.exists(new_path):
os.rename(old_path, new_path)
print(f"Renamed: {file} β†’ {new_file}")
files_renamed += 1
print(f"Total files renamed: {files_renamed}")
def load_audio_files(folder):
"""Load all WAV files from a folder into memory"""
files = sorted([f for f in os.listdir(folder) if f.endswith(".wav")])
stems = {}
for file in tqdm(files, desc="Loading audio files"):
path = os.path.join(folder, file)
audio = AudioSegment.from_wav(path)
stems[file] = audio
return stems
def get_stems(folder):
"""Get a list of all WAV files in a folder"""
files = sorted([f for f in os.listdir(folder) if f.endswith(".wav")])
return files
def apply_audio_operation(audio, operation, value):
"""Apply various audio operations to an AudioSegment"""
if operation == "low_pass_filter" and isinstance(value, float):
return low_pass_filter(audio, value)
elif operation == "high_pass_filter" and isinstance(value, float):
return high_pass_filter(audio, value)
elif operation == "fade_in" and isinstance(value, int):
return audio.fade_in(value)
elif operation == "fade_out" and isinstance(value, int):
return audio.fade_out(value)
elif operation == "reverb" and isinstance(value, float):
# Simple reverb simulation by adding delayed and attenuated copies
result = audio
for delay in [50, 100, 150, 200]:
attenuated = audio - (value * 10) # Reduce volume based on reverb value
delayed = AudioSegment.silent(duration=delay) + attenuated
result = result.overlay(delayed)
return result
elif operation == "delay" and isinstance(value, int):
# Simulate delay by adding a delayed copy
result = audio
delayed = AudioSegment.silent(duration=value) + (audio - 6) # -6dB for the echo
return result.overlay(delayed)
elif operation == "distortion" and isinstance(value, float):
# Simulate distortion by adding some limiting/clipping
gain = 1.0 + (value * 5) # Boost the gain based on distortion value
return audio + (gain) # Add gain in dB
elif operation == "pitch_shift" and isinstance(value, float):
# Note: pydub doesn't natively support pitch shifting
print(f"Warning: Pitch shift not implemented, value: {value}")
return audio
elif operation == "volume" and isinstance(value, float):
# Adjust volume by dB
return audio + value
return audio
def create_section_from_json(section_config, stems):
"""Create an audio section based on JSON configuration"""
if not section_config:
print("No configuration found for section")
return AudioSegment.empty()
section_stems = []
print(section_config)
for stem_name in section_config["stems"]:
# First try the original stem name
if stem_name in stems:
section_stems.append(stems[stem_name])
else:
# Try the name without spaces
no_spaces_name = stem_name.replace(" ", "")
if no_spaces_name in stems:
section_stems.append(stems[no_spaces_name])
else:
print(f"Warning: Stem {stem_name} not found (with or without spaces)")
# Apply operations to stems
processed_stems = {name: audio for name, audio in stems.items()}
for op in section_config.get("operations", []):
stem_name = op["stem"]
operation = op["operation"]
value = op["value"]
# Check both original and no-spaces versions
stem_key = None
if stem_name in processed_stems:
stem_key = stem_name
else:
no_spaces_name = stem_name.replace(" ", "")
if no_spaces_name in processed_stems:
stem_key = no_spaces_name
if stem_key and operation != "overlay":
processed_stems[stem_key] = apply_audio_operation(
processed_stems[stem_key], operation, value
)
# Collect the processed stems for this section
final_stems = []
for stem_name in section_config["stems"]:
if stem_name in processed_stems:
final_stems.append(processed_stems[stem_name])
else:
no_spaces_name = stem_name.replace(" ", "")
if no_spaces_name in processed_stems:
final_stems.append(processed_stems[no_spaces_name])
# Overlay stems if specified
# if section_config.get("overlay", True) and final_stems:
result = final_stems[0]
for stem in final_stems[1:]:
result = result.overlay(stem)
# return result
# Remove silences longer than 1.5 seconds (1500 ms)
silence_thresh = result.dBFS - 16 #
silent_chunks = silence.detect_silence(
result, min_silence_len=1500, silence_thresh=silence_thresh
)
segments = []
prev_end = 0
for start, end in silent_chunks:
if prev_end < start:
segments.append(result[prev_end:start])
prev_end = end
segments.append(result[prev_end:])
result = sum(segments)
return result
return AudioSegment.empty()
def generate_section_variants(
stems_folder, audio_stems, section_type, bpm, bars,progress, p=0.5
):
"""
Generate multiple variants for a specific section
Args:
stems_folder (str): Path to folder containing stem files
section_type (str): Type of section (intro, verse, chorus, etc.)
bpm (int): Beats per minute
bars (int): Number of bars
p (float): Variation parameter (0-1)
Returns:
dict: Dictionary of variant audio segments and their descriptions
"""
stems = get_stems(stems_folder)
try:
llm_response = make_groq_call(
stems,
f"{section_type} section",
p,
section_type=section_type,
bpm=bpm,
bars=bars,
)
except Exception as e:
logger.error(f"Error generating variants for {section_type} section: {e}")
return {}
# Load audio files
if not audio_stems:
print("No stems loaded.")
return {}
progress(0.2, desc=f"Generating structure and effects for variants of {section_type}...")
# Create each variant
# hERE IS tHE validation for the LLM RESPONSE
if not isinstance(llm_response, dict):
logger.error(f"Invalid LLM response: {llm_response}")
# Check if the response is a valid JSON object
try:
if isinstance(llm_response, str):
llm_response = json.loads(llm_response)
elif isinstance(llm_response, dict):
llm_response = llm_response
except json.JSONDecodeError:
logger.error(f"Tried Converting the LLM response to JSON, but failed: {llm_response}")
variants = {}
for variant_key in progress.tqdm(llm_response, desc="Getting variants as per AI arranegments..."):
if variant_key.startswith("variant"):
variant_config = llm_response[variant_key]
logger.info(f" gsv: Variant config: {variant_config}")
audio = create_section_from_json(variant_config, audio_stems)
description = variant_config.get(
"description", f"Variant {variant_key[-1]}"
)
list_stems = variant_config.get("stems", [])
variants[variant_key] = {
"audio": audio,
"description": description,
"config": variant_config,
"stems": list_stems
}
return variants
def create_full_track(
sections_folder, audio_stems, selected_variants, crossfade_ms=500
):
"""
Create a full track from selected variants
Args:
sections_folder (dict): Dict mapping section names to their folder paths
selected_variants (dict): Dict mapping section names to their selected variant configs
crossfade_ms (int): Crossfade duration in milliseconds
Returns:
AudioSegment: The final track
"""
final_track = None
# Define the order of sections
section_order = [
"intro",
"buildup",
"full_loop",
"breakdown",
"bridge",
"buildup2",
"drop2",
"breakdown2",
"outro",
]
# Process each section in order
for section_name in section_order:
if section_name not in selected_variants:
continue
# Get the selected variant config
variant_config = selected_variants[section_name]
# Create audio for this section
section_audio = create_section_from_json(variant_config, audio_stems)
# Add to final track
if final_track is None:
final_track = section_audio
else:
final_track = final_track.append(section_audio, crossfade=crossfade_ms)
return final_track
def create_intro(llm_answer, stems):
"""Create intro section from LLM answer"""
return create_section_from_json(llm_answer.get("create_intro", {}), stems)
def create_variation1(llm_answer, stems):
"""Create variation1 section from LLM answer"""
return create_section_from_json(llm_answer.get("create_variation1", {}), stems)
def create_full_loop(llm_answer, stems):
"""Create full loop section from LLM answer"""
return create_section_from_json(llm_answer.get("create_full_loop", {}), stems)
def create_variation2(llm_answer, stems):
"""Create variation2 section from LLM answer"""
return create_section_from_json(llm_answer.get("create_variation2", {}), stems)
def create_variation3(llm_answer, stems):
"""Create variation3 section from LLM answer"""
return create_section_from_json(llm_answer.get("create_variation3", {}), stems)
def create_outro(llm_answer, stems):
"""Create outro section from LLM answer"""
return create_section_from_json(llm_answer.get("create_outro", {}), stems)
def calculate_duration(bpm, bars):
"""Calculate duration in seconds for a given BPM and number of bars"""
# Assuming 4/4 time signature (4 beats per bar)
beats_per_bar = 4
duration_seconds = (bars * beats_per_bar * 60) / bpm
return duration_seconds
def get_formatted_duration(seconds):
"""Format duration in seconds to MM:SS format"""
minutes = int(seconds // 60)
seconds = int(seconds % 60)
return f"{minutes}:{seconds:02d}"
def export_section_variants(variants, output_folder, section_name):
"""Export section variants to audio files"""
if not os.path.exists(output_folder):
os.makedirs(output_folder)
file_paths = {}
for variant_key, variant_data in variants.items():
output_path = os.path.join(output_folder, f"{section_name}_{variant_key}.wav")
variant_data["audio"].export(output_path, format="wav")
file_paths[variant_key] = output_path
return file_paths
def edm_arrangement_tab():
with gr.Blocks() as iface:
gr.Markdown("# 🌚 Interactive EDM Arrangement Tool")
with gr.Row():
variation = gr.Radio(
choices=list(arrangements.keys()),
value="High Energy Flow",
label="Choose Arrangement Variation",
)
out_plot = gr.Plot(label="Arrangement Diagram", value=load_variation("High Energy Flow"))
variation.change(fn=load_variation, inputs=variation, outputs=out_plot)
iface.load(fn=load_variation, inputs=[variation], outputs=[out_plot])
with gr.Accordion("🎻 Edit Section Parameters", open=False):
for i, (bar, tempo, name, length, curve) in enumerate(arrangement):
with gr.Row():
gr.Markdown(f"**{name}**")
bar_slider = gr.Slider(
minimum=0, maximum=300, value=bar, label="Start Bar"
)
tempo_slider = gr.Slider(
minimum=20, maximum=100, value=tempo, label="Volume"
)
length_slider = gr.Slider(
minimum=1, maximum=64, value=length, label="Length"
)
curve_selector = gr.Radio(
choices=["Flat", "Linear", "-ve Linear"],
value=curve,
label="Curve Type",
)
update_btn = gr.Button("Update")
update_btn.click(
fn=update_section,
inputs=[
gr.Number(value=i, visible=False),
bar_slider,
tempo_slider,
length_slider,
curve_selector,
],
outputs=[out_plot],
)
with gr.Accordion("βž• Insert New Section", open=False):
new_index = gr.Number(value=0, label="Insert At Index")
new_name = gr.Textbox(label="Section Name", value="New Section")
new_bar = gr.Slider(minimum=0, maximum=300, value=0, label="Start Bar")
new_tempo = gr.Slider(minimum=20, maximum=100, value=50, label="Volume")
new_length = gr.Slider(minimum=1, maximum=64, value=8, label="Length")
new_curve = gr.Radio(
choices=["Flat", "Linear", "-ve Linear"], value="Flat", label="Curve Type"
)
insert_btn = gr.Button("Insert Section")
insert_btn.click(
fn=insert_section,
inputs=[new_index, new_name, new_bar, new_tempo, new_length, new_curve],
outputs=[out_plot],
)
gr.Markdown("## βœ… Finalise Your Arrangement")
final_btn = gr.Button("Finalise and Export JSON")
final_output = gr.Textbox(label="Final Arrangement JSON", lines=15)
final_btn.click(fn=finalise, outputs=final_output)