MCD & F0

​

Function Signature

mcd_f0(
    pred_x,
    gt_x,
    fs,
    f0min,
    f0max,
    mcep_shift=5,
    mcep_fftl=1024,
    mcep_dim=39,
    mcep_alpha=0.466,
    seq_mismatch_tolerance=0.1,
    power_threshold=-20,
    dtw=False
)

​

Parameters

​

Returns

​

Usage Examples

​

Basic Usage (No DTW)

import numpy as np
from versa import mcd_f0

# Load audio signals
reference = np.random.random(16000)  # Replace with actual reference
generated = np.random.random(16000)  # Replace with actual generated audio
fs = 16000

# Calculate MCD and F0 metrics
results = mcd_f0(
    pred_x=generated,
    gt_x=reference,
    fs=fs,
    f0min=80,   # Male voice range
    f0max=400,
    dtw=False
)

print(f"MCD: {results['mcd']:.2f} dB")
print(f"F0 RMSE: {results['f0rmse']:.2f} Hz")
print(f"F0 Correlation: {results['f0corr']:.3f}")

​

With Dynamic Time Warping

import numpy as np
from versa import mcd_f0

# For sequences with timing differences
reference = np.random.random(24000)
generated = np.random.random(20000)  # Different length OK with DTW
fs = 16000

results = mcd_f0(
    pred_x=generated,
    gt_x=reference,
    fs=fs,
    f0min=100,  # Female voice range
    f0max=600,
    dtw=True,   # Enable DTW alignment
    power_threshold=-20
)

print(f"MCD (DTW): {results['mcd']:.2f} dB")
print(f"F0 RMSE: {results['f0rmse']:.2f} Hz")
print(f"F0 Correlation: {results['f0corr']:.3f}")

​

Custom MCEP Configuration

import numpy as np
from versa import mcd_f0

reference = np.random.random(22050)
generated = np.random.random(22050)
fs = 22050

# Custom configuration for 22.05 kHz
results = mcd_f0(
    pred_x=generated,
    gt_x=reference,
    fs=fs,
    f0min=80,
    f0max=500,
    mcep_shift=5,
    mcep_fftl=2048,      # Larger FFT for higher SR
    mcep_dim=39,
    mcep_alpha=0.410,    # Alpha for 22.05 kHz
    dtw=False
)

print(f"MCD: {results['mcd']:.2f} dB")

​

Technical Details

​

Feature Extraction Pipeline

1. Preprocessing
   • Scale audio to int16 range
   • Apply low-cut filter (70 Hz cutoff)
   • Handle multi-channel by using first channel
2. WORLD Vocoder Analysis
   • F0 extraction using HARVEST algorithm
   • Spectral envelope via CheapTrick
   • Aperiodicity estimation with D4C
   • Convert spectral envelope to mel-cepstrum
3. MCD Calculation
   • Compute frame-wise Euclidean distance
   • Apply conversion factor: 10/ln(10) * sqrt(2 * sum(diff^2))
   • Average across all frames
4. F0 Metrics
   • Extract only voiced frames (F0 > 0)
   • Calculate RMSE and correlation

​

DTW vs Non-DTW

• Direct frame-by-frame comparison
• Requires similar sequence lengths (within tolerance)
• Faster computation
• Best for well-aligned signals

• Applies Voice Activity Detection
• Aligns sequences using Dynamic Time Warping
• Handles timing variations
• Power-based alignment for MCD
• F0-based alignment for F0 metrics

​

Parameter Guidelines

​

F0 Range by Voice Type

​

Alpha Values by Sampling Rate

​

FFT Length Recommendations

​

Interpretation

​

MCD Values

• < 4.5 dB: Excellent similarity (professional quality)
• 4.5 - 6.5 dB: Good similarity (acceptable quality)
• 6.5 - 8.5 dB: Moderate similarity (noticeable differences)
• > 8.5 dB: Poor similarity (significant degradation)

​

F0 Correlation

• > 0.9: Excellent pitch contour matching
• 0.7 - 0.9: Good pitch tracking
• 0.5 - 0.7: Moderate correlation
• < 0.5: Poor pitch matching

​

Dependencies

pip install pyworld pysptk scipy fastdtw librosa

​

Use Cases

• Voice conversion quality evaluation
• Speech synthesis assessment
• Singing voice synthesis validation
• Speaker adaptation evaluation
• Prosody transfer measurement

​

Error Handling