How AI Diagnoses Sound System Problems

SonaVyx uses two complementary diagnostic approaches. Tier 1 is a rule-based engine that runs instantly in the browser, checking measurement data against known problem patterns. It detects feedback (narrow spectral peaks), hum (50/60 Hz harmonics), polarity errors (cross-correlation sign), comb filtering (periodic nulls), clipping (flat-top waveforms), noise floor issues, and harmonic distortion. This tier is free and always available.

Tier 2 uses the Claude API (Sonnet 4) for deep contextual analysis. The measurement data is sent to Claude along with a specialized system prompt containing acoustic engineering knowledge. Claude analyzes the data holistically, considering how different parameters interact, and generates a comprehensive diagnostic report with specific recommendations.

The rule-based engine checks six categories. Frequency balance: is the response within plus or minus 6 dB across the 100 Hz to 10 kHz range? Phase alignment: are there comb filter patterns indicating timing misalignment? RT60: is reverberation within the target range for the room type? STI: is speech intelligibility above the threshold? SPL: is the noise floor acceptable? Coherence: is the measurement reliable?

Each check produces a pass, warning, or fail result with specific details (frequency, magnitude, severity). The combined results generate a health score from 0 to 100 that summarizes overall system performance at a glance.

When you request a full AI diagnostic, SonaVyx packages the measurement data (frequency response, phase, coherence, RT60, SPL, and any problem detector results) and sends it to Claude via the Anthropic API. The system prompt instructs Claude to act as an acoustic consultant, analyzing the data for specific problem patterns and generating structured recommendations.

Claude returns a JSON response with a health score, category scores (frequency balance, phase alignment, coverage, noise floor, reverb, equipment match), a prioritized problem list with severity levels, and specific EQ and treatment recommendations. The structured format ensures consistent, actionable output.

A typical AI diagnostic session starts with a measurement capture. The user runs the transfer function or RTA for 30 seconds, then clicks "Diagnose." The rule engine produces instant results, identifying any obvious problems. If the user wants deeper analysis, they click "Full Diagnostic" which sends data to Claude.

Within 5-10 seconds, the AI returns a comprehensive report. Example output: "Health Score: 62/100. Primary issues: 1) 8 dB peak at 250 Hz suggesting room mode excitation (reduce via 4 dB parametric cut at 250 Hz, Q=2). 2) RT60 of 1.8s exceeds target of 0.8s for this room type (ceiling treatment recommended). 3) Coherence drops below 0.5 above 8 kHz (possible measurement noise or speaker HF rolloff)."

All DSP processing runs locally in the browser. Your audio never leaves your device. Only when you explicitly request an AI diagnostic does measurement data (frequency response numbers, not audio recordings) leave the device. The Claude API processes the data and returns the analysis. No audio is stored on SonaVyx servers or by the API provider. The analysis results are stored in your account if you are logged in, or discarded if you are not.