DTW Audio Comparison — User Guide

Dynamic Time Warping analysis comparing recordings of different lengths/tempos, measuring MFCC similarity, tempo-aligned loudness, and melodic interval accuracy with automatic quality assessment.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel

What this does

This script uses Dynamic Time Warping (DTW) to compare two audio recordings that may have different durations or tempos. Unlike basic comparison tools, DTW intelligently aligns sequences to handle timing variations, making it ideal for comparing performances with different speeds, rhythmic variations, or duration mismatches. Analyzes MFCC features (timbre), loudness, and melodic intervals with tempo-aware alignment.

What is DTW? Dynamic Time Warping is an algorithm that finds optimal alignment between two time sequences. Think of it like stretching/compressing a rubber band to match two patterns. It's widely used in speech recognition, gesture recognition, and music analysis to compare signals that don't line up perfectly in time.

Quick start

In Praat, select exactly two Sound objects (Shift+click to select both).
Run script… → Load this DTW comparison script.
Click OK — processing completes in 10-30 seconds (depends on duration).
Check Info window for detailed results.
Check Objects window for "dtw_analysis_results" table.
Optional: Save table as CSV (Select table → Save → Save as comma-separated values file).

Key advantage: Works with different-length recordings! Teacher recording can be 10s, student 12s — DTW handles it. Ideal for comparing performances at different tempos (student slower/faster than teacher).

Three-Dimensional Analysis

🎨 DTW MFCC Similarity (Timbre)

What it measures: Spectral/timbral similarity using Mel-Frequency Cepstral Coefficients

Interpretation:

Normalized DTW distance <2: Excellent (very similar timbre)
2-5: Good (comparable tone quality)
5-10: Moderate (noticeable timbral differences)
>10: Needs work (significantly different tone)

Use: Assesses tone quality, articulation, vowel shape (voice), instrument timbre matching

🔊 DTW Loudness Match (Dynamics)

What it measures: Tempo-aligned intensity comparison

Interpretation:

Aligned dB difference <4: Excellent (similar dynamic control)
4-8 dB: Good (comparable loudness patterns)
8-15 dB: Moderate (noticeable dynamic differences)
>15 dB: Needs work (different dynamic approach)

Use: Evaluates dynamic expression, loudness control matching

🎵 DTW Pitch Accuracy (Melody)

What it measures: Tempo-aligned melodic interval comparison

Interpretation:

Aligned interval difference <0.3 semitones: Excellent
0.3-0.7 semitones: Very good
0.7-1.5 semitones: Good
1.5-3.0 semitones: Moderate
>3.0 semitones: Needs work

Additional metrics:

Contour accuracy: Percentage of matched melodic direction (up/down/stable)
Transposition: Overall pitch difference (e.g., singing in wrong key)

DTW Advantages

Handles tempo variations:

Student performs faster/slower than teacher → DTW aligns them
Rubato or rhythmic variations → Flexible alignment
Different durations (10s vs 15s) → Compares intelligently

Timing flexibility:

No need for exact time alignment
Accounts for accelerando/ritardando
Finds best correspondence between sequences

vs. Standard comparison:

Standard: Requires same duration, frame-by-frame comparison, fails with tempo changes
DTW: Adaptive alignment, robust to timing differences, finds optimal matching

Important: DTW works best when recordings have similar content (same piece/exercise). Comparing completely different material gives meaningless results. Recordings should start at approximately the same point (beginning of phrase/piece).

Output Files

Info window displays:

Duration difference and tempo ratio
DTW MFCC, Loudness, and Pitch assessments
Contour accuracy percentage
Overall transposition (key difference)

Results Table (dtw_analysis_results):

Column	Description
parameter	Feature type (MFCC_DTW, LOUDNESS_DTW, PITCH_DTW)
metric	Specific measurement name
value	Numeric result
assessment	Quality rating (EXCELLENT/GOOD/MODERATE/NEEDS_WORK)
teacher_file	Reference recording name
student_file	Test recording name
duration_diff	Duration difference (seconds)
tempo_ratio	Speed ratio (student/teacher)

To save results: Select "dtw_analysis_results" table → Save → Save as comma-separated values file → Choose location

Parameters

MFCC Parameters (timbre analysis):

Parameter	Value	Description
number_of_filters	24	Mel filter bank size
fmin/fmax	100-5000 Hz	Frequency range for analysis
number_of_coefficients	13	MFCC features (C0-C12, using C1-C12)

DTW Parameters:

Parameter	Value	Description
dtw_radius	3	Search radius for alignment
dtw_min_size	5	Minimum sequence length for FastDTW

Pitch Parameters:

Parameter	Value	Description
pitch_floor/ceiling	75-600 Hz	Range for F0 detection

Note: All parameters hardcoded in script, optimized for speech/singing analysis.

Applications

Music pedagogy:

Compare student performance to teacher demonstration (different tempos OK)
Assess timbre matching, dynamic expression, pitch accuracy simultaneously
Track improvement across multiple recordings at different speeds
Evaluate rubato/expressive timing variations

Language learning:

Compare native speaker to learner (natural speech rate differences)
Assess pronunciation (MFCC), intonation (pitch), and prosody (loudness)
Handle hesitations and timing differences in speech

Performance analysis:

Compare multiple takes at different tempos
Analyze tempo consistency across performances
Identify specific measures/phrases with issues (via frame analysis)

Limitations

Processing time: 10-30 seconds for typical recordings (DTW computationally intensive)
Similar content required: Recordings must be same piece/exercise (not random material)
Approximate alignment: Uses FastDTW approximation for long sequences (sampling every 3rd frame)
Start point matters: Recordings should begin at approximately same musical point
Monophonic preferred: Pitch analysis works best with single melodic line
Large tempo differences: Ratio >2.0× may reduce accuracy

Interpreting Results

Example 1: Excellent match

MFCC: 1.5 (EXCELLENT) → Same tone quality
Loudness: 3.2 dB (EXCELLENT) → Similar dynamics
Pitch: 0.2 semitones (EXCELLENT) → Accurate intonation
Contour: 95% → Correct melodic direction
Tempo ratio: 1.15 → Student 15% slower (acceptable)

Example 2: Needs work

MFCC: 12.0 (NEEDS_WORK) → Different timbre/articulation
Loudness: 18 dB (NEEDS_WORK) → Poor dynamic control
Pitch: 3.5 semitones (NEEDS_WORK) → Significant intonation errors
Contour: 60% → Wrong melodic direction often
Transposition: 7 semitones → Wrong key entirely

Example 3: Mixed results (common)

MFCC: 4.0 (GOOD) → Decent tone quality
Loudness: 11 dB (MODERATE) → Dynamics need attention
Pitch: 0.8 semitones (GOOD) → Minor intonation issues
Action: Focus practice on dynamic control