TechnoTaggr

Audio tagging toolkit built on Essentia TensorFlow models using the essentia-tensorflow package. TechnoTaggr scans folders of audio, runs prepackaged classifiers (mood, tonality, instruments, loops, Jamendo genres, etc.), and saves structured JSON that you can post-process into 16‑bar phrase summaries and explore through a Plotly Dash UI.

Here is a demo of the Dashboard with Ploty

In progress Work:

Data Analysis:

• Data analysis usint initial aggregation logic for 16 bar measures common in techno music. Creating data set of low-level audio features from mel-spectrograms. Establishing baseline reference of accuracy for high-level features from classifiers on curated techno music dataset.

• Curating new techno dataset + finetuning models

• Adding functionality to add tags into wav, mp3, and .aiff files.

What’s inside

• Prebundled Essentia feature extractors (MusiCNN, Discogs EffNet) and classification heads in src/technotaggr/models
• CLI for batch analysis (analyze), 16‑bar phrase aggregation (postprocess), and interactive visualization (visualize)
• JSON session logging with per-segment probabilities and aggregated scores
• Dash dashboard with dark theme, per-model plots, and session metadata panels
• Python API for embedding extraction, classification, and result logging

Setup (uv)

uv venv
source .venv/bin/activate
uv sync

# Optional: refresh bundled Essentia models
uv run python src/technotaggr/download_models.py

CLI usage

Run commands with uv run (or directly if the package is installed). Supported audio formats: .mp3, .wav, .aiff.

Analyze audio

Scan a directory, run every discovered classifier, and write a timestamped results file to technotaggr_results/.

uv run technotaggr analyze song_data/smol \
  --output-dir technotaggr_results \
  --recursive \
  --verbose

• Discovers audio files, loads matching feature extractors/classifiers, and caches embeddings per file to avoid redundant work.
• Each classifier yields per-segment probabilities plus an aggregated mean per class.
• A session summary prints to the console and is saved as results_YYYYMMDD_HHMMSS.json.

Post-process 16‑bar phrases

Estimate BPM with librosa, compute 16‑bar phrase (bar) probabilities from the segment outputs, and append them to the results JSON.

uv run technotaggr postprocess technotaggr_results/results_*.json \
  --audio-base-path song_data/smol \
  --output technotaggr_results/results_postprocessed.json

• Adds bpm, phrase_duration_seconds, per-phrase bar_predictions, and aggregated_bar_predictions (mean across phrases).
• Uses embedding model patch sizes (with 50% overlap) to infer segment duration; falls back to audio-length/segment-count if needed.
• Prints a per-embedding summary of segment duration, segments per phrase, and number of phrases found.

Visualize results (Dash)

Launch the Dash dashboard to explore a session file (defaults to the latest JSON in technotaggr_results).

uv run technotaggr visualize \
  --session-file technotaggr_results/results_postprocessed.json \
  --port 8050

• Session dropdown: pick any results_*.json.
• Audio dropdown: pick a track within that session.
• Info panel: shows session timestamp, input/output folders, total/success/fail counts, and classifiers used.
• For each model:
  • Line plot of per-segment probabilities (always)
  • Line plot of 16‑bar phrase probabilities (when post-processed data is present)
  • Side-by-side bar charts of aggregated segment means and aggregated phrase means

Result file shape

Results live in technotaggr_results/ as JSON:

• Session metadata: session_timestamp, input_directory, output_directory, total_files, successful_files, failed_files, classifiers_used
• results: one entry per audio file with audio_file, audio_duration_seconds, sample_rate, and models
  • Each model contains model_name, model_version, paths to classifier/embedding graphs, classes, num_segments, segment_predictions, and aggregated_predictions
  • After postprocess, models also include bar_predictions and aggregated_bar_predictions, and the audio entry gains bpm and phrase_duration_seconds

Python API quickstart

from pathlib import Path
from technotaggr import InferencePipeline, discover_classifiers

classifiers = discover_classifiers()  # loads configs from src/technotaggr/models
pipeline = InferencePipeline(classifiers)
result = pipeline.analyze_audio(Path("song.mp3"))

for pred in result.predictions:
    top_class, score = max(pred.aggregated_predictions.items(), key=lambda x: x[1])
    print(pred.classifier_name, top_class, score)

Models

• Classification heads included: loop detection, tonal/atonal, mood (aggressive/happy/relaxed/sad), NSynth instrument/reverb, Jamendo genre, and more under models/classification-heads/.
• Feature extractors included: MusiCNN (msd-musicnn-1) and Discogs EffNet (discogs-effnet-bs64-1).
• If you need to re-fetch them, run uv run python src/technotaggr/download_models.py (uses curl to pull from Essentia’s model hub).

Back to top