EXPERIMENT 02 · LOCAL-FIRST · 12 NOVEMBER 2026

Music Probev0.4

Turn a digital recording into evidence an assistant can inspect.

Measure the signal. Keep interpretation honest. This downloadable Python prototype distinguishes audio from video streams and analyzes an excerpt on your own computer. It does not listen like a human or identify instruments, lyrics, performers or emotion.

Download v0.4 · Python ZIPSHA-256 checksum

Four views of the same signal

Stereo evidence

Separate left/right waveforms and spectra, channel levels, correlation and mid/side energy. True two-channel sources only; no invented stereo.

Rhythm over time

Onset and energy traces at 0.1 or 0.01 second steps, with competing tempo candidates. Tempo uses six-second windows, not ten milliseconds of sound.

Spectrogram

Frequency versus time, useful for spotting bands and transients. Resampled to 22.05 kHz, so the display stops near 11 kHz.

Shareable reports

JSON and text, plus CSV and optional PNG plots. A video soundtrack can be analyzed, but video frames are not interpreted.

Run it locally

Unzip and open the included README. Requires Python 3.10+, NumPy, and current FFmpeg/ffprobe. Install matplotlib separately for plots. Use trusted official packages and only media you are entitled to process.

python -m pip install -r requirements.txt
python -m pip install matplotlib
python music_probe.py "clip.mp4" --stereo \
  --spectrogram --rhythm-step 0.1 \
  --output analysis.json

Omit --stereo for mono audio. Default excerpt: 60 seconds; maximum: 300 seconds. Existing outputs are protected; choose a new report name for another run.

What the numbers cannot tell you

Correlation is not perceived stereo width. Repetition is not necessarily a beat. A spectrogram cannot establish genre or meaning. Base rhythm measurements use mono downmix; separate stereo measurements preserve both channels. Resampling and overlapping windows limit precision.

The optional remote interpretation route is separate from local measurement. It uploads an extracted mono clip to Google only with both explicit remote flags and an API key. Provider terms, processing rights, retention and costs still apply. It is not required for this release.

Prototype validation

Sixteen offline tests cover synthetic tones, silence, short input, pulse timing, stream detection, output protection, upload gating and stereo cases. Synthetic tests check engineering behavior, not musical intelligence. Run python -m unittest -v inside the extracted directory.

Experimental software, not a forensic, medical, copyright-identification or professional loudness-certification tool. No private recordings, API credentials or personal reports are bundled.