AI Acoustic Classification

AI Acoustic Classification turns recorded sound into a decision: which class does this sample belong to, and how confident is the model? The complete chain — annotating samples, training a model, running inference, and rolling the model out to a line station — runs inside Deepsona. No external analysis engine is required, and no measurement data has to leave your machine.

Typical uses are end-of-line pass/fail judgement (OK vs. whine vs. knock vs. rub), sorting a large archive of recordings by fault type, and screening for abnormal units when defective samples are too rare to label.

The workflow at a glance

The four stages form a loop: every batch you classify produces new material you can annotate and feed back into training, so the model improves as the product line accumulates data.

Opening the workspace

Open the Start Menu and choose AI Classification. The workspace is a step-by-step wizard; the step indicator across the top shows where you are, and the arrow buttons on the right move between steps. You may jump directly to any step — for example, if you already have a trained model, go straight from Load Model to Classify and skip training entirely.

Step	What it does	Needed when
1. Load Model	Select the `.onnx` file and optionally enable GPU	Always
2. Train Model	Turn a labelled folder into a model	Only if you are building your own model
3. Settings	Class labels and score post-processing	Recommended before the first run
4. Classify	Run one file and inspect all class scores	Verification and single-sample work
5. Batch Classify	Run a whole file list and export CSV	Archive sorting, line reporting

Step 1 — Load Model

Browse to an ONNX model file. Deepsona reads the model's input and output signature and shows a short summary (for example Input: [1,1,16000] -> Output: [1,10] (10 classes)), so you can immediately confirm that the file is the model you intended and how many classes it predicts. Tick GPU (DirectML) to use the graphics card; if no compatible device is present the engine falls back to the CPU without an error.

Details of shape matching, activation options and the pre-trained Model Zoo are covered in ONNX Inference.

Step 2 — Train Model

Two training routes are offered side by side. ML.NET (zero-code) trains entirely inside Deepsona — you point it at a folder of labelled audio and press Start. Python (deep learning) hands the job to a PyTorch script; two ready-made templates (a CNN classifier and an autoencoder) can be exported with one click and customised. Both routes finish with a model file you load in Step 1.

See Annotation & Training for the folder layout, the data volume you need, and how to read the training metrics.

Step 3 — Settings

Enter the class labels as a comma-separated list, in the same order the model outputs them (for example OK, Whine, Knock, Rub). Without labels, results are reported as Class_0, Class_1, and so on — the prediction is still correct, only harder to read.

The post-processing options (Sigmoid, SoftMax, Winner Takes All, Temperature, Top-K) convert the model's raw output numbers into readable confidences. Defaults work for most models; the individual meanings are described in ONNX Inference.

Step 4 — Classify

Browse to a WAV or HEAD HDF file, or hand data over from the workspace, and press Run. The headline result shows the predicted class and its confidence, and below it every class is listed with a bar and a percentage — the winning class is shown in bold. Always look at the full list rather than the headline alone: two classes at 45 % and 43 % mean something very different from one class at 98 %.

Step 5 — Batch Classify

Add any number of WAV or HDF files, press Run Batch, and watch the progress bar. Each row of the table shows the file name, predicted label, confidence and status; failed files are marked individually instead of aborting the run. Export CSV writes the whole table (File, PredictedLabel, Confidence, Status) for use in a report or a quality database. Cancel stops a long run cleanly, keeping the rows already computed.

What data do you need?

File formats — WAV and HEAD HDF are read directly. Multi-channel files are handled; models that expect a single channel receive a mono mixdown automatically.
Consistency matters more than quantity — use the same microphone position, the same operating point and the same recording length for training and for production. A model trained on 3-second run-up recordings should be fed 3-second run-up recordings.
Balanced classes — aim for a comparable number of samples per class. If one class dominates heavily, the model tends to answer with it regardless of the input.
A held-back set — keep some labelled recordings out of training and use Batch Classify on them; that is the honest measure of how the model will behave on the line.

Typical scenarios

Scenario	Recommended approach
End-of-line pass/fail with several known fault types	Label samples per fault type, train a classifier, deploy with Batch Classify
Sorting a historical archive by fault type	Load an existing model, run Batch Classify, export CSV and sort
Defects are rare or unknown in advance	Anomaly detection — train on normal samples only
Few labelled samples, many unlabelled recordings	Start with a pre-trained Model Zoo model or embeddings, then refine as labels accumulate

An AI result is an indication, not a measurement. Confidence values describe how strongly the model prefers one class over the others — they are not an error bound and not a physical quantity. For release decisions, pair the classification with the objective analyses (level, loudness, order spectra) that the rest of Deepsona provides.

Where to go next

ONNX Inference — model formats, shape matching, GPU, activations, Model Zoo
Annotation & Training — data layout, ML.NET and Python routes
Anomaly Detection — autoencoders and statistical baselines
AI Assistant & Reporting — turning results into documents