Audio Feature Analysis for STT Benchmark

Overview

This analysis examines correlations between audio characteristics and Word Error Rate (WER) to understand why certain microphones perform better for speech-to-text transcription.

Key Findings

Top Correlations with WER

• Spectral Bandwidth Mean: r=-0.516 (p=0.049)

  • higher values correlate with FEWER errors
  • statistically significant

• Harmonic Ratio: r=-0.443 (p=0.098)

  • higher values correlate with FEWER errors
  • not statistically significant

• Speech Clarity Ratio: r=0.393 (p=0.147)

  • higher values correlate with MORE errors
  • not statistically significant

• Mfcc 2 Mean: r=-0.336 (p=0.221)

  • higher values correlate with FEWER errors
  • not statistically significant

• Speech Formant Ratio: r=0.325 (p=0.237)

  • higher values correlate with MORE errors
  • not statistically significant

• Spectral Rolloff Mean: r=-0.313 (p=0.255)

  • higher values correlate with FEWER errors
  • not statistically significant

• Speech Fundamental Ratio: r=-0.259 (p=0.352)

  • higher values correlate with FEWER errors
  • not statistically significant

• Rms Mean: r=0.244 (p=0.381)

  • higher values correlate with MORE errors
  • not statistically significant

• Spectral Centroid Mean: r=-0.242 (p=0.386)

  • higher values correlate with FEWER errors
  • not statistically significant

• Zero Crossing Rate Mean: r=0.166 (p=0.555)

  • higher values correlate with MORE errors
  • not statistically significant

Category Analysis

Category	Avg WER	WER Std	Spectral Centroid	Harmonic Ratio	Formant Energy
Desktop	5.31%	0.60%	2439 Hz	0.551	50.5%
Headset	5.93%	0.18%	2507 Hz	0.297	62.9%
Lavalier	5.40%	nan%	2302 Hz	0.460	53.9%
Mobile	5.32%	0.83%	2187 Hz	0.517	63.6%

Best and Worst Performers

Best (Lowest WER)

• Sample 15 (OnePlus Nord 3 5G): 4.13% WER
  • Spectral Centroid: 2639 Hz, Harmonic Ratio: 0.532
• Sample 1 (UGreen CM564): 4.44% WER
  • Spectral Centroid: 2893 Hz, Harmonic Ratio: 0.548
• Sample 7 (Audio-Technica ATR4697): 4.76% WER
  • Spectral Centroid: 2850 Hz, Harmonic Ratio: 0.667

Worst (Highest WER)

• Sample 13 (Audio-Technica ATR4750-USB): 6.35% WER
  • Spectral Centroid: 2555 Hz, Harmonic Ratio: 0.485
• Sample 11 (Yealink BH72): 6.03% WER
  • Spectral Centroid: 1526 Hz, Harmonic Ratio: 0.413
• Sample 12 (Yealink BH72): 6.03% WER
  • Spectral Centroid: 1532 Hz, Harmonic Ratio: 0.366

Interpretation

The analysis reveals several counterintuitive findings:

1. Audio quality scores don't predict STT accuracy: Traditional measures of audio quality (SNR, dynamic range) show weak correlation with transcription accuracy.

2. Smartphone processing helps: The OnePlus phone applies real-time audio processing (noise reduction, AGC) that appears to benefit STT engines, even though it reduces "audiophile" quality metrics.

3. Spectral characteristics matter: Features like spectral centroid and harmonic ratio show meaningful correlations, suggesting that frequency balance and voice clarity impact recognition accuracy.

4. Environment > Equipment: The same microphone (OnePlus Nord 3) shows dramatically different WER in quiet (4.13%) vs noisy (6.03%) environments, confirming that recording conditions dominate equipment choice.

Note on Statistical Significance

With only 15 samples, many correlations don't reach traditional statistical significance (p < 0.05). These findings are indicative rather than conclusive. A larger sample size would be needed to confirm these patterns.