📦 View on PyPI

Note: Formerly known as amt-augpy. Starting with v1.0.9, the package is amt-augmentor.

A Python toolkit for augmenting Automatic Music Transcription (AMT) datasets through conventional audio transformations while maintaining synchronization between audio and MIDI files. Its optional CSV helpers follow the same metadata concepts as MAESTRO v3.0.0.

The toolkit expects a folder containing paired audio and MIDI files with matching names. The versioned paired Python APIs accept MIDI containing multiple annotated instruments. The legacy batch CLI converts MIDI through a note-only intermediate representation, so it does not retain instrument or non-note event structure. The pair must already be aligned ground truth: this toolkit augments existing datasets; it does not create annotations from unlabelled audio.

dataset/
├── song1.wav        # Audio file
├── song1.mid        # Ground truth annotated midi file

Features

Audio Transformations

• Time Stretching: Tempo modification while maintaining pitch
• Pitch Shifting: Transposition while preserving timing
• Reverb & Filtering: Room acoustics and frequency filtering effects
• Gain & Chorus: Depth and richness enhancement
• Noise Augmentation: Controlled noise addition for robustness training
• Fractional Detuning: Audio detuning by less than 50 cents with unchanged symbolic note labels
• Archival Noise (opt-in): Configurable seeded 1/f-like noise and harmonic mains hum for datasets whose target recordings have those characteristics
• Legacy Audio Merging: Disabled by default because sources must be assigned to splits before any merge can be shown to be leakage-safe

Processing & Dataset Handling

• Audio Standardization: Channel-preserving, non-destructive conversion to 44.1 kHz WAV working copies
• Parallel Processing: Multi-core processing for faster augmentation
• Configuration System: YAML-based parameter customization
• Dataset Validation: Automatic validation of train/test/validation splits
• Dataset Modification: Built-in tools to modify existing dataset splits
• MAESTRO Compatibility: Dataset format compatible with MAESTRO v3.0.0

Why AMT-Augmentor?

Built for AMT, not just audio. Unlike general audio augmenters, the versioned paired APIs keep audio and MIDI aligned through transform-consistent label updates: pitch shift transposes pitched notes, while time-changing methods map note boundaries, control changes, pitch bends, and supported global timed events. The toolkit also provides MAESTRO-style CSV construction and split validation.

Deterministic paired transforms

The versioned Python API provides deterministic paired implementations of gain/chorus, target-SNR noise, reverb and filtering, integral and fractional pitch changes, global time stretch, and local time variation. Each transform validates the audio--MIDI pair, records its parameters and synchronization checks, and publishes source-bound provenance. Useful strengths and parameter ranges depend on the instrument, recording conditions, annotation policy, and downstream model; no research recipe is enabled implicitly.

Several methods were refined while studying historical Hardanger fiddle recordings, but they are exposed as configurable AMT primitives rather than as a fiddle-specific pipeline:

• reverb_only_v1 separates room simulation from high-pass and low-pass filtering so their effects can be evaluated independently.
• materialize_pitch_shift_grid_v1 renders and verifies an explicit, caller-supplied integral pitch grid. Audio and MIDI are transposed together, and callers may set narrower output pitch bounds for a particular model. Drum tracks are rejected because General MIDI drum numbers identify kit pieces rather than transposable pitches.
• fractional_detuning_v1 changes audio by less than half a semitone while leaving symbolic MIDI pitches unchanged.
• archival_noise_v1 models configurable coloured noise and mains hum. It is useful only when those conditions resemble the intended target domain.
• local_time_warp_v1 renders a complete recording once through a continuous, monotonic time map and applies that exact map to MIDI event boundaries. It does not cut, repeat, or splice chunks.

These methods are opt-in. Their effect on transcription accuracy should be measured on held-out source groups for each dataset rather than assumed from another instrument or corpus.

from amt_augmentor import NoiseSNRParameters, noise_snr_v1

noise_snr_v1(
    "tune.wav",
    "tune.mid",
    "tune_augmented_noise.wav",
    "tune_augmented_noise.mid",
    seed=42,
    parameters=NoiseSNRParameters(target_snr_db=24.0),
)

from amt_augmentor import (
    ArchivalNoiseParameters,
    FractionalDetuningParameters,
    archival_noise_v1,
    fractional_detuning_v1,
)

fractional_detuning_v1(
    "tune.wav",
    "tune.mid",
    "tune_detuned.wav",
    "tune_detuned.mid",
    seed=42,
    parameters=FractionalDetuningParameters(cents=30.0),
)

archival_noise_v1(
    "tune.wav",
    "tune.mid",
    "tune_archival.wav",
    "tune_archival.mid",
    seed=42,
    parameters=ArchivalNoiseParameters(
        target_snr_db=28.0,
        hum_power_fraction=0.20,
        mains_frequency_hz=50.0,  # Use the value appropriate for the corpus.
        harmonic_count=3,
    ),
)

For an integral pitch study, pass the tested shifts explicitly rather than relying on a package-wide grid:

from amt_augmentor import materialize_pitch_shift_grid_v1

materialize_pitch_shift_grid_v1(
    "tune.wav",
    "tune.mid",
    "pitch_grid",
    seed=42,
    semitones=(-2, -1, 1, 2),
)

The research-facing functions require a nonnegative integer seed, validate all selected parameters, and record them in provenance. They refuse to overwrite source or output files, validate the audio/MIDI pair (including relevant note bounds), preserve channel count, and write a JSON sidecar containing source and output hashes plus the exact parameter plan. Time-stretched and pitch-shifted labels are serialized as high-resolution, constant-tempo AMT annotations; they are not intended as symbolic scores with an inherited tempo map.

The audio, MIDI, and provenance files form one logical bundle. Payloads are staged first, the provenance sidecar is published last as the completion marker, and caught failures are rolled back. Consumers should only accept a bundle when the sidecar exists and its hashes match.

Mixed-audio note removal, experimental synthetic pause insertion, and an experiment-specific composite materializer were withdrawn after review and are not part of the package API. Their prototypes remain available through Git history only. The historical add_pause implementation is retained solely to reproduce old runs because the legacy CLI depends on it. It is disabled by default, omitted from the advertised effects, emits a visible warning when explicitly enabled, and must not be used for new training data. See docs/AUGMENTATION_SCOPE.md for the decision and current evaluation scope.

Research case study

A configuration evaluated on historical Hardanger fiddle recordings is described separately in the Hardanger fiddle case study. It supports the manuscript Automatic Music Transcription for Oral Traditions: Methodology and a Hardanger Fiddle Case Study (in preparation). The case study documents one corpus-specific use of the toolbox; it is neither a package default nor a claim that the same settings are optimal for other AMT datasets. The exact Galdr experiment manifests and training adapters belong with that research artifact, not in the general-purpose package API.

Requirements

• Python 3.9, 3.10, 3.11, 3.12, or 3.13
• System dependencies: libsndfile and ffmpeg (for audio processing)

Installation

You can install AMT-Augmentor either via pip or by cloning the repository:

Using pip

pip install amt-augmentor

From source

git clone https://github.com/LarsMonstad/amt-augmentor.git
cd amt-augmentor
pip install -e .

Usage

Basic Usage

amt-augmentor /path/to/dataset/directory
# Or running directly
python -m amt_augmentor.main /path/to/dataset/directory

This processes compatible audio files and their corresponding MIDI files using the legacy batch configuration. That workflow samples parameters from the YAML ranges; the deterministic paired Python APIs shown above instead use explicit, provenance-recorded settings. The legacy workflow flattens notes into a simple event annotation during processing, so use the paired Python APIs when MIDI instrument tracks, controls, or pitch bends must be retained.

Advanced Usage

# Use a custom configuration file
amt-augmentor /path/to/dataset/directory --config my_config.yaml

# Set random seed for reproducible augmentation
# (forces num_workers=1 — worker subprocesses don't inherit RNG state)
amt-augmentor /path/to/dataset/directory --seed 42

# Reproducible train/test/validation split (independent of --seed)
amt-augmentor /path/to/dataset/directory --split-seed 7

# Specify an output directory
amt-augmentor /path/to/dataset/directory --output-directory /path/to/output

# Generate a default configuration file
amt-augmentor --generate-config my_config.yaml

# Disable specific effects
amt-augmentor /path/to/dataset/directory --disable-effect timestretch --disable-effect chorus

# Control legacy merge behavior via the YAML config (merge_audio.merge_num)
# (no CLI flag — see config.sample.yaml for the merge_audio.merge_num key)

# Modify existing dataset CSV files
amt-augmentor --modify-csv dataset.csv --list-split all  # List all songs
amt-augmentor --modify-csv dataset.csv --move-to-split test --song-patterns "Mozart"  # Move songs
amt-augmentor --modify-csv dataset.csv --remove-songs --song-patterns "BadRecording"  # Remove songs

# Parallel processing with 8 workers
amt-augmentor /path/to/dataset/directory --num-workers 8

# Custom train/test/validation split
amt-augmentor /path/to/dataset/directory --train-ratio 0.8 --test-ratio 0.1 --validation-ratio 0.1

# Force specific songs to test set (prevents augmentation)
amt-augmentor /path/to/dataset/directory --custom-test-songs "song1,song3,song5"

# Force specific songs to validation set (prevents augmentation)
amt-augmentor /path/to/dataset/directory --custom-validation-songs "song2,song4"

# Dry run to preview what will be processed
amt-augmentor /path/to/dataset/directory --dry-run

# Verbose output for debugging
amt-augmentor /path/to/dataset/directory --verbose

# Check for valid MIDI-WAV pairs before processing
amt-augmentor /path/to/dataset/directory --check-pairs

# List available effects
amt-augmentor --list-effects

# Check version
amt-augmentor --version

Help and options

amt-augmentor --help

Configuration

All augmentation parameters can be customized using a YAML configuration file. See config.sample.yaml for a complete example with documentation.

File Format Support

Audio

• Input: WAV, FLAC, MP3, M4A, AIFF
• Output: WAV (44.1kHz)

Annotations

• MIDI (.mid)

Output Structure

On first run, the toolkit reorganizes your dataset into two subfolders:

<dataset>/
    original/    # your pristine audio + MIDI pairs (moved from the input dir)
    augmented/   # every augmented file produced by the pipeline

Keeping augmented output in its own folder means you can delete augmented/ to reset the dataset without touching any source material.

Augmented files follow the naming convention:

original_name_augmented_effect_parameter_randomsuffix.extension

The _augmented_ identifier ensures all augmented files are properly recognized and handled during dataset creation. Example of augmented/ contents:

piano_augmented_timestretch_1.2_abc123.wav
piano_augmented_timestretch_1.2_abc123.mid
piano_augmented_noise_1.5_def456.wav
piano_augmented_noise_1.5_def456.mid

The generated CSV references files with their subfolder prefix (<dataset>/original/... and <dataset>/augmented/...), so the physical layout mirrors the CSV exactly.

Dataset Creation & Validation

The dataset follows the same format as MAESTRO v3.0.0. Songs assigned to test or validation splits will have their augmented versions excluded to prevent data leakage.

Creating the Dataset CSV

# Create dataset with default split ratios (70% train, 15% test, 15% validation)
amt-augmentor /path/to/directory

# Create dataset with custom split ratios
amt-augmentor /path/to/directory --train-ratio 0.8 --test-ratio 0.1 --validation-ratio 0.1

# Force specific songs to test set (they won't be augmented)
amt-augmentor /path/to/directory --custom-test-songs "song1,song3,song5"

# Force specific songs to validation set (they won't be augmented)
amt-augmentor /path/to/directory --custom-validation-songs "song2,song4"

--custom-test-songs and --custom-validation-songs use case-insensitive substring matching against the song stem. If the same title matches both lists, test wins and a warning is printed. Pinned songs are also skipped at augmentation time — no augmented WAVs/MIDIs are written for them — so held-out evaluation data stays untouched on disk and in the CSV. Substring matching means --custom-test-songs "piece_17" will pin every variant (piece_17_take1, piece_17_take2, piece_17_studio, ...) to the same split.

Validating the Dataset Split

Dataset split validation is automatically performed after CSV creation to ensure:

• Augmented songs are not included in test/validation splits
• No cross-split contamination occurs (an augmented row's source original must live in the same split)
• Every augmented row has a matching original (no "orphan aug" rows)

You can also run validation as a standalone, side-effect-free check — useful after hand-editing a CSV, merging datasets, or receiving a split from someone else:

# Human-readable report (exits 0 regardless)
amt-augmentor --validate-csv dataset.csv

# CI-friendly: non-zero exit when contamination is found
amt-augmentor --validate-csv dataset.csv --strict

# Machine-readable JSON (for piping into other tooling)
amt-augmentor --validate-csv dataset.csv --json

# Equivalent direct invocation of the validator module
python -m amt_augmentor.validate_split dataset.csv --strict

CSV Format

The generated CSV follows the MAESTRO format with the following columns:

• canonical_composer
• canonical_title
• split
• year
• midi_filename
• audio_filename
• duration

Modifying Existing Datasets

After creating a dataset CSV, you can easily modify it to adjust train/test/validation splits:

# List all songs and their distribution
amt-augmentor --modify-csv dataset.csv --list-split all

# List only test songs
amt-augmentor --modify-csv dataset.csv --list-split test

# List all songs with detailed view
amt-augmentor --modify-csv dataset.csv --list-split all --verbose

# Move songs to a different split (substring matching)
amt-augmentor --modify-csv dataset.csv --move-to-split test --song-patterns "Mozart,Chopin"

# Remove songs from dataset
amt-augmentor --modify-csv dataset.csv --remove-songs --song-patterns "BadRecording1,BadRecording2"

# Create backup before modifications (off by default)
amt-augmentor --modify-csv dataset.csv --move-to-split validation --song-patterns "Bach" --backup

Features:

• Substring matching: Patterns like "Mozart" match any song containing that substring
• Smart augmented handling: Augmented versions automatically stay in train split only
• Backup option: Use --backup to create a backup before modifications

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

For development:

1. Install development dependencies: pip install -e ".[dev]"
2. Run tests: pytest tests/
3. Check typing: mypy amt_augmentor
4. Format code: black amt_augmentor

Contributors

• Lars Monstad (@LarsMonstad) – Original author and maintainer
• @monoamine11231 – Noise augmentation, custom test songs feature, and various improvements

Contact

For questions or collaboration:

• Email: lars@botsformusic.com
• Organization: https://botsformusic.com
• GitHub: https://github.com/LarsMonstad/amt-augmentor

License

MIT License - see LICENSE file for details.

Citation

If you use this toolkit in your research, please cite:

@software{amt_augmentor,
  author       = {Lars Monstad and contributors},
  title        = {AMT-Augmentor: Audio + MIDI augmentation toolkit for AMT datasets},
  version      = {2.0.0},
  year         = {2026},
  publisher    = {Bots for Music},
  url          = {https://github.com/LarsMonstad/amt-augmentor}
}