Dolby A Dataset

As a result, there is a large amount of historical audio currently stored on tape that is encoded in the Dolby A format. These tapes require the complementary decoding process to be applied for digital transfer. In order to facilitate the development of models that can reproduce that process digitally so that these artifacts might be preserved, we present a dataset of music and synthetic probe signals processed by two revisions of Dolby A hardware.</p><p dir="ltr">## Dataset Summary</p><p dir="ltr">* **6132** total mono input examples</p><p dir="ltr">* **3056** music input clips</p><p dir="ltr">* **3056** synthetic training probe input examples</p><p dir="ltr">* **21** synthetic analysis probe input examples</p><p dir="ltr">* **2** hardware revisions</p><p dir="ltr">* **12264** hardware encode examples in the main metadata table</p><p dir="ltr">* **612** hardware encode/decode round-trip reference examples</p><p dir="ltr">* **9** predefined subsets, each with train, validation, and test splits</p><p><br></p><p dir="ltr">* Sample rate: **44.1 kHz**</p><p dir="ltr">* Resolution: **24-bit PCM**</p><p dir="ltr">## Source Audio</p><p dir="ltr">All music inputs were sourced from MoisesDB, which contains the multitracks for 240 songs.

To prepare the dataset, the full mixes were reconstructed by summing associated individual source files. The left and right channels of each mix were treated as independent mono sources. The original source files were then collapsed to mono.

From each source, a 30-second excerpt was chosen pseudorandomly, and a 10 ms fade was applied at each end. Because many of the individual tracks are quite sparse, a -60 dBFS RMS threshold was implemented to ensure that the excerpted portion contained meaningful content. If that threshold was not met, a new random excerpt was chosen.

This process was repeated with a patience of 100, at which point the file was skipped. A -3 dBFS peak ceiling was imposed on all excerpts; if the excerpt was found to exceed that ceiling, it was normalized to -3 dBFS peak.</p><p dir="ltr">The music subset contains 3056 30-second input clips: 480 full-mix excerpts and 2576 individual-source excerpts.</p><p dir="ltr">To generate the training probes, we multiply an MLS carrier by a piecewise-constant random amplitude envelope whose hold durations are sampled log-uniformly and whose active levels are sampled uniformly in decibels over a bounded range.

Each hold is independently assigned to be active or inactive. Inactive holds are set to digital silence, while active holds are assigned levels between -54 dBFS and -12 dBFS. Hold durations range from 20 ms to 3 s, producing both rapid level transitions and longer intervals over which the Dolby A ballistics can approach steady state.</p><p dir="ltr">The dataset also includes 21 synthetic analysis probes for diagnostic use.

These include sine ramps, sine bursts, an MLS ramp, MLS bursts, and a silence probe. They are included as auxiliary analysis material, but the predefined training subsets use the music and synthetic training-probe corpora.</p><p dir="ltr">## Contents</p><p dir="ltr">The dataset includes:</p><p dir="ltr">* Mono input audio signals</p><p dir="ltr">* Corresponding Dolby A hardware encode captures</p><p dir="ltr">* A small set of hardware encode/decode round-trip reference captures</p><p dir="ltr">* Dolby tone calibration files</p><p dir="ltr">* A small set of synthetic analysis probes and process-isolation captures</p><p dir="ltr">* Metadata describing source material, probe parameters, hardware revision, split membership, and signal relationships</p><p dir="ltr">## Data Organization</p><p dir="ltr">All relationships between signals are defined explicitly in metadata tables.

File paths in the metadata tables are relative to the repository root.</p><p dir="ltr">### Primary table</p><p dir="ltr">* `Metadata/metadata_main.parquet`</p><p dir="ltr">Defines all hardware encode pairs and provides paths to:</p><p dir="ltr">* input audio</p><p dir="ltr">* output audio</p><p dir="ltr">Each row corresponds to a single unprocessed input and Dolby A hardware-encoded output. For decoder modeling, the encoded output is typically used as the model input and the unprocessed input is used as the target.</p><p dir="ltr">### Supporting tables</p><p dir="ltr">* `Metadata/metadata_music_inputs.parquet`</p><p dir="ltr">Music input metadata, including MoisesDB source identifiers, track information, stem names, source type, channel role, excerpt position, and level statistics</p><p dir="ltr">* `Metadata/metadata_training_probe_inputs.parquet`</p><p dir="ltr">Synthetic training-probe metadata, including MLS-envelope generation parameters, random seeds, duration, and level statistics</p><p dir="ltr">* `Metadata/metadata_analysis_probes.parquet`</p><p dir="ltr">Synthetic analysis-probe metadata, including probe family, frequency, burst, ramp, or silence parameters, duration, level statistics, and auxiliary analysis capture paths</p><p dir="ltr">* `Metadata/metadata_decode_test_split_music-362_363.parquet`</p><p dir="ltr">Hardware encode/decode round-trip reference metadata for the music 362/363 test set</p><p dir="ltr">* `Metadata/metadata_decode_test_split_probes-362_363.parquet`</p><p dir="ltr">Hardware encode/decode round-trip reference metadata for the probes 362/363 test set</p><p dir="ltr">### File structure</p><p dir="ltr">* `Inputs/Music/` — music input clips</p><p dir="ltr">* `Inputs/TrainingProbes/` — synthetic training probe input clips</p><p dir="ltr">* `Inputs/AnalysisProbes/` — synthetic analysis probe input clips</p><p dir="ltr">* `Outputs/Encoded/Music/` — Dolby A encoded music captures</p><p dir="ltr">* `Outputs/Encoded/TrainingProbes/` — Dolby A encoded training-probe captures</p><p dir="ltr">* `Outputs/Encoded/AnalysisProbes/` — Dolby A encoded analysis-probe captures</p><p dir="ltr">* `Outputs/Decoded/Music/` — hardware encode/decode round-trip reference captures for the music 362/363 test set</p><p dir="ltr">* `Outputs/Decoded/TrainingProbes/` — hardware encode/decode round-trip reference captures for the probes 362/363 test set</p><p dir="ltr">* `Outputs/Decoded/AnalysisProbes/` — process-isolation captures made by passing analysis probes directly through the hardware decode path</p><p dir="ltr">* `Outputs/DA->AD_only/AnalysisProbes/` — DA/AD-only analysis-probe captures</p><p dir="ltr">* `Outputs/Calibration/` — Dolby tone calibration files</p><p dir="ltr">* `Metadata/` — parquet tables</p><p dir="ltr">The filesystem layout is not required to infer relationships between signals.

All mappings are defined in metadata.</p><p dir="ltr">### Hardware</p><p dir="ltr">While all Dolby A processors perform the same primary function, there are multiple hardware implementations that realize that functionality using various revisions of the circuit.</p><p dir="ltr">These differences can lead to slight variations in the magnitude response, noise, and total harmonic distortion across models.</p><p dir="ltr">Dolby A hardware is generally implemented as a mainframe with one or more plug-in cards on which most of the noise-reduction circuitry resides.</p><p dir="ltr">For this dataset, we employed two two-channel mainframes, Models 362 and 363.</p><p dir="ltr">The 362 frame was fitted with two CAT No. 22 cards, while the 363 housed a pair of CAT No. 300 modules.</p><p dir="ltr">Both the 362 frame and CAT No. 22 modules are constructed from discrete transistor circuitry, and the frame employs transformer-balanced inputs and outputs.</p><p dir="ltr">The later 363 and CAT No. 300 designs rely more heavily on integrated circuits, and the frame I/O is electronically balanced.</p><p dir="ltr">The inclusion of two distinct circuit revisions allows for models to be optimized to emulate specific devices or for revision-agnostic models to be trained using data from both devices.</p><p dir="ltr">All input signals are processed by both units.</p><p dir="ltr">### Predefined subsets</p><p dir="ltr">The primary Hugging Face interface is a set of nine predefined dataset subsets.

Each subset is represented by three parquet files:</p><p dir="ltr">* `Metadata/Subsets//train.parquet`</p><p dir="ltr">* `Metadata/Subsets//validation.parquet`</p><p dir="ltr">* `Metadata/Subsets//test.parquet`</p><p><br></p><p dir="ltr">CSV mirrors of each split are included beside the parquet files for manual inspection. The parquet files are the canonical files used by the Hugging Face dataset loader.</p><p dir="ltr">The subsets are:</p><p dir="ltr">| Subset | Input condition | Hardware condition |</p><p dir="ltr">| `music-362` | Music only | 362 only |</p><p dir="ltr">| `music-363` | Music only | 363 only |</p><p dir="ltr">| `music-362_363` | Music only | 362 and 363 |</p><p dir="ltr">| `probes-362` | Training probes only | 362 only |</p><p dir="ltr">| `probes-363` | Training probes only | 363 only |</p><p dir="ltr">| `probes-362_363` | Training probes only | 362 and 363 |</p><p dir="ltr">| `music_probes-362` | Music and training probes | 362 only |</p><p dir="ltr">| `music_probes-363` | Music and training probes | 363 only |</p><p dir="ltr">| `music_probes-362_363` | Music and training probes | 362 and 363 |</p><p dir="ltr">Input IDs are assigned to train, validation, and test splits before subsets are constructed.

These assignments are preserved across all subsets so that an input signal used for testing in one subset does not appear as training or validation data in another.</p><p dir="ltr">## Calibration Signals</p><p dir="ltr">Dolby tone calibration signals are included in `Outputs/Calibration/`. The hardware units used to create the dataset were calibrated to a Dolby Level of +4 dBu, and the included Dolby tone files provide a reference for that calibration.</p><p dir="ltr">Analysis probes are also included to support inspection of:</p><p dir="ltr">* Pseudo-static response using ramp signals</p><p dir="ltr">* Attack and release behavior using burst signals</p><p dir="ltr">* Frequency-dependent behavior at representative Dolby A band regions</p><p dir="ltr">* MLS-based diagnostic behavior</p><p dir="ltr">* Direct encode-path, decode-path, and DA/AD-only behavior</p><p dir="ltr">## Temporal Alignment</p><p dir="ltr">System latency was measured and compensated for during capture so that input and output signals are aligned in time to within ~1 sample.</p><p dir="ltr">## Important Notes</p><p dir="ltr">* File paths in the metadata are relative to the repository root</p><p dir="ltr">* All audio files are mono WAV files sampled at 44.1 kHz</p><p dir="ltr">* `audio_in_path` points to the unprocessed source signal</p><p dir="ltr">* `audio_out_path` points to the corresponding hardware-processed signal</p><p dir="ltr">* The main supervised decoder target is the original unprocessed source signal, not the hardware-decoded reference output</p><p dir="ltr">* Music and training-probe decode files are encode/decode round-trip reference captures</p><p dir="ltr">* Analysis-probe decode files are direct decode-path captures intended to isolate that hardware process</p><p dir="ltr">* Analysis-probe DA/AD-only files are included to isolate the converter path from the Dolby hardware path</p><p dir="ltr">* Calibration tones and analysis probes are auxiliary files and are not part of the predefined train, validation, or test splits</p><p dir="ltr">* The music portion is derived from MoisesDB; users should ensure that downstream use complies with the relevant source-material terms.</p><p dir="ltr">## Intended Use</p><p dir="ltr">This dataset is intended for:</p><p dir="ltr">* Black-box modeling of Dolby A encode and decode behavior</p><p dir="ltr">* Development and evaluation of software Dolby A decoders for archival audio transfer</p><p dir="ltr">* System-identification experiments for nonlinear, dynamic audio processors</p><p dir="ltr">* Cross-evaluation of models trained on music, synthetic probes, or both</p><p dir="ltr">* Hardware-specific, pooled-hardware, and hardware-holdout modeling experiments</p><p dir="ltr">The dataset is not intended to model arbitrary tape restoration problems or abnormal hardware operation.

It does not attempt to characterize overload, clipping, tape noise, wow and flutter, azimuth error, medium defects, or deliberate analog saturation.</p><p dir="ltr">The music portion is derived from MoisesDB. Users should ensure that downstream use complies with the terms associated with the source material as well as the license for this dataset.</p><p dir="ltr">## Citation</p><p dir="ltr">If you use this dataset, please cite the dataset and accompanying paper.</p><p><br></p><p dir="ltr">### Dataset</p><p dir="ltr">@dataset{thompson2026dolbyadataset,</p><p dir="ltr">title = {{Dolby A Dataset}},</p><p dir="ltr">author = {Thompson, Benjamin R. and Chitsakul, Pakaphol and Heilemann, Michael C.},</p><p dir="ltr">year = {2026},</p><p dir="ltr">url = {huggingface.co/datasets/bthomp23/Dolby_A_Dataset},</p><p>}</p><p><br></p><p><br></p><p dir="ltr">### Paper</p><p><br></p><p dir="ltr">@inproceedings{thompson2026dolbyacasestudy,</p><p dir="ltr">title = {{Dolby A: A Dataset and Case Study in Black-Box Modeling of a Nonlinear Audio Processor}},</p><p dir="ltr">author = {Thompson, Benjamin R. and Chitsakul, Pakaphol and Heilemann, Michael C.},</p><p dir="ltr">booktitle = {Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing},</p><p dir="ltr">year = {2026},</p><p>}</p><p><br></p>