Audio Analysis

Audio & Music Analysis Skill (Powered by SonicBridge v0.6.0)

1. Overview

This skill guides AI agents in performing deep signal-processing analysis on local music tracks using the high-performance sonic-bridge Rust engine.

Instead of relying on heavy pre-trained models or complex external Python libraries, the agent invokes sonic-bridge to decouple raw waveforms into LRMD (LLM-Readable Music Descriptor) reports. This allows pure-text LLMs to "listen to" and "appreciate" the dynamic tempo, timbral changes, and chord progressions of any cover version or original master track with millisecond-level precision.

Key Features (v0.6.0)

• Multi-Detector Ensemble Architecture: 5 specialized detectors (Ambient, Chinese Pentatonic, Western Classical, Pop/Electronic, Jazz Rubato) with confidence-weighted voting
• OPIH Time-Signature Engine: Automatic 3/4 waltz detection and BPM correction
• Pentatonic Tonic-Share Scaling: Accurate detection of Chinese/Japanese pentatonic modes (宫商角徵羽)
• Confidence Scoring: Joint confidence metric combining BPM and Key detection reliability
• Style Classification: Automatic genre detection (Pop/Rock/Electronic, Classical, Chinese Folk/Modal, Jazz/Rubato, Ambient/Free)

2. Technical Analysis via SonicBridge CLI

When a track is downloaded locally via alx download <id>, the agent should call the sonic-bridge CLI tool to parse its acoustic structure.

Command-Line Usage

# 1. Standard Spatiotemporal Analysis (Default 5s step)
sonic-bridge "/path/to/song.mp3"

# 2. Beat-Synchronous Analysis (Recommended for most music)
sonic-bridge "/path/to/song.mp3" --beat

# 3. Event-Driven Onset Adaptive Segmentation (Best for fast/complex tracks)
sonic-bridge "/path/to/fast_melody_song.mp3" --onset

# 4. Silent Mode (Only generates .lrmd.md, no stdout preview)
sonic-bridge "/path/to/song.mp3" --beat --quiet

# 5. Cross-Version Comparative Analysis (DTW Aligner)
sonic-bridge "/path/to/original.mp3" "/path/to/cover_version.mp3"

PITFALL: Always use --quiet when analyzing multiple files in batch to avoid flooding stdout. The .lrmd.md file is always generated regardless of --quiet.

3. LRMD Protocol Specifications

The sonic-bridge tool automatically generates a <filename>.lrmd.md report in the same directory. The agent must read this file to interpret the musical metadata.

Example LRMD Structure (v0.6.0)

# SonicBridge: LLM-Readable Music Descriptor (LRMD)

## 1. Global Acoustic & Musicological Metadata
- **Filename**: `The Weeknd - Blinding Lights.flac`
- **Duration**: `201.57 seconds`
- **Tempo (BPM)**: `172.3 BPM` (Extremely Rapid (Presto))
- **Estimated Key**: `F Minor`
- **Primary Style**: `Pop/Rock/Electronic`
- **Analysis Confidence**: `0.63`

## 2. Spatiotemporal Track Analysis (Adaptive Onset Intervals)
| Timeline | Chord | Dynamic Intensity | Timbral Brightness | Rhythmic & Transient Activity |
| :--- | :--- | :--- | :--- | :--- |
| **0.0s - 1.0s** | `Dm` | Very Soft (Pianissimo) | Deep & Dark (Muddy/Sub-heavy) | Flowing & Legato (Gentle melodic flow) |
| **1.0s - 1.1s** | `Dsus2` | Soft & Intimate (Piano) | Deep & Dark (Muddy/Sub-heavy) | Flowing & Legato (Gentle melodic flow) |
| **1.1s - 4.8s** | `Fm` | Soft & Intimate (Piano) | Deep & Dark (Muddy/Sub-heavy) | Flowing & Legato (Gentle melodic flow) |

New Fields in v0.6.0

Confidence Score Interpretation

PITFALL: When confidence < 0.45, the BPM/Key results may be unreliable. Consider flagging these tracks for manual review in automated pipelines.

4. Distribution Analysis Patterns

The LRMD report enables three types of distribution analysis:

Dynamic Distribution

grep -oP "(Silent|Very Soft|Soft & Intimate|Moderately Intense|Loud & Energetic|Loud & Dense)" file.lrmd.md | sort | uniq -c | sort -rn

Timbral Brightness Distribution

grep -oP "(Balanced & Clear|Warm & Smooth|Bright & Crisp|Deep & Dark|Piercing & Airy)" file.lrmd.md | sort | uniq -c | sort -rn

Rhythmic Activity Distribution

grep -oP "(Static & Ambient|Flowing & Legato|Steady Beat|Walking Groove|Driving & Punchy|Syncopated)" file.lrmd.md | sort | uniq -c | sort -rn

5. Agent Companionship & Conversation Patterns

By interpreting the LRMD report, the Agent can provide connoisseur-level music companionship and emotional alignment.

Conversation Pattern 1: Multi-Version Comparative Critique

When the user plays a cover version of a song, the agent can call process_comparative and discuss the aesthetic difference:

• Agent Dialogue Prompt:

  "Compared to Eason Chan's original version which relies on a lush, wet reverb space (RT60 ~2.4s) to build cinematic gravity, the acoustic cover you are playing right now is ultra-minimalistic. The single acoustic guitar has dry, close-mic transient attacks, and the singer's voice features heavy breathiness (Airy Timbre), creating a profoundly intimate, heartbreaking kitchen-table conversation vibe."

Conversation Pattern 2: Causal Harmonic Resolution Guidance

Explain how the musical tension resolves to comfort the user's mood:

• Agent Dialogue Prompt:

  "I noticed that at the 0.6s mark of this intro, the arrangement performs a rapid harmonic shift from A Major to A Minor, which resolves into F Major at 0.7s. That fleeting subdominant-to-tonic tension release is precisely why this song feels so comforting yet melancholic."

Conversation Pattern 3: Style-Aware Recommendation

Use the Primary Style and Confidence fields to make informed recommendations:

• Agent Dialogue Prompt:

  "This track has been classified as 'Traditional Chinese Folk/Modal' with 0.67 confidence. The SonicBridge analysis detected pentatonic scale patterns (E 商调式) — this means the melody avoids the 4th and 7th degrees, creating that distinctive Eastern modal sound. If you enjoy this, I can find other tracks with similar pentatonic characteristics."

Conversation Pattern 4: Confidence-Based Quality Assessment

Use the confidence score to assess analysis reliability:

• Agent Dialogue Prompt:

  "The analysis confidence for this track is only 0.39 — this indicates the arrangement is highly experimental or non-standard. The BPM and Key results should be taken with a grain of salt. This is typical for avant-garde or free improvisation music."

6. Batch Analysis Workflow

For analyzing large music collections:

# 1. Analyze all files in a directory
find /mnt/data/Music/Jazz/ -name "*.flac" -o -name "*.mp3" | while read f; do
  sonic-bridge "$f" --beat --quiet
done

# 2. Extract metadata to CSV for analysis
find /mnt/data/Music/ -name "*.lrmd.md" | while read f; do
  bpm=$(grep "Tempo" "$f" | grep -oP '\d+\.\d+')
  key=$(grep "Estimated Key" "$f" | grep -oP '`[^`]+`' | tr -d '`')
  style=$(grep "Primary Style" "$f" | cut -d'`' -f2)
  conf=$(grep "Analysis Confidence" "$f" | grep -oP '0\.\d+')
  echo "$(basename "$f"),$bpm,$key,$style,$conf"
done > /tmp/music_metadata.csv

# 3. Filter by confidence threshold
awk -F',' '$5 >= 0.55' /tmp/music_metadata.csv > /tmp/high_confidence_tracks.csv

PITFALL: When batch analyzing 100+ files, use --quiet to avoid stdout flooding. Monitor progress by checking the count of .lrmd.md files generated.