Python3中scikits.talkbox的lpc无法使用，求替代方案及共振峰频率估算示例

Ah, I feel your pain—scikits.talkbox is basically abandoned for Python 3, so that LPC function is definitely not reliable these days. Let’s walk through some solid replacements and a complete example to estimate formants in Python 3 that actually works.

1. Use Librosa's LPC Implementation #

Librosa is a well-maintained, widely used audio processing library that fully supports Python 3. Its librosa.lpc function is a drop-in replacement for the old scikits version.

First, install Librosa if you haven't:

pip install librosa

Then use it like this:

import librosa

# y = your audio signal array, n_order = desired LPC order
lpc_coeffs = librosa.lpc(y, order=n_order)

2. Use python_speech_features (Lightweight Option) #

If you prefer a library focused specifically on speech features, python_speech_features also has a reliable LPC implementation:

Install it first:

pip install python_speech_features

Call the function like so:

from python_speech_features import lpc

# signal = your audio signal, numcep = LPC order
lpc_coeffs = lpc(signal, numcep=n_order)

Formants are the resonant frequencies of the vocal tract, and we can calculate them from LPC coefficients by finding roots of the LPC polynomial. Here's a full, working example using Librosa:

import librosa
import numpy as np

def estimate_formants(y, sr, lpc_order=12):
    # Optional: Apply pre-emphasis to boost high frequencies (reduces noise impact)
    y_preemphasized = librosa.effects.preemphasis(y)
    
    # Extract LPC coefficients
    lpc_coeffs = librosa.lpc(y_preemphasized, order=lpc_order)
    
    # Find roots of the LPC polynomial
    roots = np.roots(lpc_coeffs)
    # Keep only roots in the upper half of the complex plane (correspond to positive frequencies)
    roots = [root for root in roots if np.imag(root) > 0]
    
    # Convert roots to frequencies (Hz)
    angles = np.arctan2(np.imag(roots), np.real(roots))
    frequencies = angles * (sr / (2 * np.pi))
    
    # Calculate magnitude of each root (indicates formant strength)
    magnitudes = 1 / np.abs(roots)
    
    # Filter out low-frequency noise (ignore values below 50Hz)
    valid_mask = frequencies > 50
    frequencies = frequencies[valid_mask]
    magnitudes = magnitudes[valid_mask]
    
    # Sort formants by frequency (and match magnitudes to the sorted order)
    sorted_indices = np.argsort(frequencies)
    frequencies = frequencies[sorted_indices]
    magnitudes = magnitudes[sorted_indices]
    
    # Return top 4 formants (F1-F4, the most useful for speech analysis)
    return frequencies[:4], magnitudes[:4]

# Example usage
# Load your audio file (replace with your file path)
y, sr = librosa.load("your_audio.wav", sr=None)

# For better accuracy, process audio in frames (this example uses the whole signal for simplicity)
formant_freqs, formant_mags = estimate_formants(y, sr)

print(f"Estimated formants (Hz): {formant_freqs.round(2)}")
print(f"Formant magnitudes: {formant_mags.round(2)}")

Quick Tips for Better Results #

• LPC Order: A good rule of thumb is 2 + (sr // 1000)—for 16kHz audio, use 18; for 8kHz, use 10. Adjust based on your specific audio.
• Frame Processing: In real-world use, split the audio into 20-30ms frames with overlap, then estimate formants per frame (Librosa's librosa.util.frame can help with this).
• Noise Reduction: If your audio is noisy, add a low-pass filter or use noise reduction techniques before estimating formants.

内容的提问来源于stack exchange，提问作者david