AI Sound Identifier

What it does

Upload an audio file or record a few seconds with your microphone, and the page lists the five sounds it thinks are present, each with a percentage and a progress bar. It uses YAMNet, a compact audio classifier trained on a broad catalogue of everyday sounds.

The output is a ranked guess about the clip as a whole, not a transcript and not a timeline. If a dog barks at the start and a person speaks at the end, you get one combined list rather than two labelled moments.

The model

YAMNet was trained on AudioSet, a large collection of short YouTube clips labelled with sound events. Its vocabulary is broad but coarse: it knows "Dog", "Speech", "Music", "Vehicle", "Alarm" and hundreds of similar categories, and it does not know the name of a song, the species of a bird beyond a general class, or the words being spoken.

The page asks the classifier for one result and then slices the top five categories. Everything below the fifth place is discarded, even if it is plausible.

Recording and uploading

There are two input paths. The round microphone button requests access to your mic and records until you press stop; a small spectrum visualiser runs while it listens. The drop zone accepts any file the browser reports as audio, either by clicking or by dragging it in.

In both cases the audio is decoded into a buffer and passed to the classifier in one call. There is no trimming, no noise reduction, no normalisation and no sample-rate control in the interface. The recording path captures whatever the browser's MediaRecorder produces, usually WebM/Opus, and decodes it back to raw samples.

If the browser cannot decode a file, the page falls back to classifying a short silent buffer instead of showing an error. A result dominated by "Silence" after an upload is a sign that the format was unsupported, not that the file was quiet.

Reading the list

Each row shows the category name on the left and the score as a percentage on the right, with a thin bar underneath. Scores are printed to one decimal place. The badge in the header says "Top 5" and does not change with the content.

The percentages are not guaranteed to sum to 100. YAMNet produces independent scores across many classes, and several can be high at once — "Speech" and "Music" often appear together in a video clip. Treat each row as a separate signal.

There is no download button and no export. The list is for reading on screen, and the status line only confirms that a sound was identified. To keep the result, copy the category names manually.

Step by step

1. Click the microphone to record, or drop an audio file onto the zone. On first use, the runtime and YAMNet weights download; the panel names each stage.
2. For a recording, press the stop button when you have captured enough; classification starts automatically.
3. For a file, decoding and classification begin as soon as it loads.
4. Read the top five categories and their percentages.
5. Try another clip or recording; the model stays loaded.

The Load AI Model button preloads the runtime if you want to avoid the first-run delay. On success it turns green and reads "Model Ready".

Where it is useful

• Nature recordings. Distinguish birdsong, insects, water and wind at a coarse level.
• Home diagnostics. Check whether a recording is dominated by an alarm, a motor, a vacuum or running water.
• Audio triage. Sort a folder of clips into speech, music and environmental sound before listening to each one.
• Content tagging. Add rough sound labels to a video library.
• Teaching. A clear demonstration of how an audio classifier turns a waveform into category scores.

It is not a Shazam-style music identifier, a speech recogniser, a species classifier or a fault detector. It gives broad categories, and the fine distinction is left to you.

Where it fails

Mixed audio is the main problem. A clip with music, speech and traffic produces high scores for all three, and the top five may omit the one sound you cared about. Because the whole clip is classified as a single input, a brief event can be averaged away by a long stretch of background noise.

Background noise and low recording levels reduce confidence across the board. So does clipping, wind on the microphone and room echo. A phone recording of a bird from ten metres away carries less usable signal than a close capture.

Fine distinctions are outside the vocabulary. The model may know "Bird" but not the species, "Musical instrument" but not which one, "Vehicle" but not the make. Genres of music, accents and individual words are not represented.

Long files are also a practical limit. There is no duration cap in the code, so a full album or a one-hour lecture will decode into a large buffer and may exhaust browser memory before classification begins.

What it does not do

No timestamps, no per-second labels, no transcript, no speaker separation, no pitch or tempo analysis, no noise removal, no trimming, no batch upload, no download and no API. It does not identify songs, recognise speech or measure loudness in decibels. It also does not tell you which part of the clip produced a label.

Privacy

Audio is decoded and classified in the browser. Nothing is uploaded, and no recording is stored after the page is closed. Microphone access is requested through the browser's own permission prompt, and the stream stops when you press stop. The runtime and model are fetched from public CDNs on first use, which reveals that the tool was loaded but not what you recorded.

Limits

One clip at a time, top five categories, no timeline, no export and no threshold control. File size and duration are limited only by browser memory. The page requires a modern browser with WebAssembly and network access on first load; microphone input also requires permission and, in most browsers, a secure page.

Common mistakes

• Expecting a transcript. The model classifies sound types, not words.
• Reading the scores as shares of the clip. They are independent confidence values.
• Uploading an unsupported format. A silent-buffer fallback can produce a meaningless "Silence" result.
• Classifying a long recording. Brief events get averaged out and memory use grows.
• Hoping for a song title. YAMNet knows "Music", not the track.
• Assuming the top label is the only sound. Several categories can be present at once.

Frequently Asked Questions