如何绘制含极小值的频谱图？librosa与matplotlib.specgram使用疑问

The Root Cause #

When you use plt.specgram(), it doesn’t just plot raw spectrum values straight up—it does a few key things under the hood that make the visualization readable, which your direct spectrum plot skips:

1. Logarithmic Amplitude Conversion: It converts linear amplitude values to decibels (dB), which compresses the huge dynamic range of audio signals. This pulls small, faint values out of the noise floor so they’re visible instead of being washed out by large peaks.
2. Automatic Dynamic Range Clipping: It typically ignores the lowest-amplitude noise (like the bottom 60-80 dB of the signal) by focusing on the meaningful range of values, which eliminates the "dark wash" from tiny, irrelevant numbers.
3. Optimized Colormap & Normalization: specgram uses colormaps (like viridis by default) tailored for spectral data, plus normalization that scales the visible range to the data’s actual meaningful values.

When you plot your raw (129, 347) spectrum array directly, you’re using a linear scale. Most of your values are near-zero noise, so they get mapped to the darkest end of your colormap—only the rare large peaks show up, making the whole plot look dim.

Fixes to Replicate specgram’s Clear Visualization #

Here’s how to adjust your custom plot to match what specgram does automatically:

1. Convert Spectrum to Decibels (dB)

Use librosa’s built-in function to compress the dynamic range—this is the most impactful fix:

import librosa
import numpy as np
import matplotlib.pyplot as plt

# Load your audio (match the parameters you used earlier)
y, sr = librosa.load("your_audio_file.wav")

# Compute spectrum with matching FFT params (129 bins = (n_fft/2)+1 → n_fft=256)
n_fft = 256
hop_length = 512  # Adjust to get your 347 time bins
D = librosa.stft(y, n_fft=n_fft, hop_length=hop_length)
spectrum_db = librosa.amplitude_to_db(np.abs(D), ref=np.max)

# Plot the dB-scaled spectrum
plt.imshow(spectrum_db, aspect="auto", origin="lower", cmap="viridis")
plt.colorbar(label="Amplitude (dB)")
plt.xlabel("Time Frames")
plt.ylabel("Frequency Bins")
plt.show()

2. Match specgram’s Exact Parameters

If you want perfect alignment with plt.specgram()’s output, capture its internal values to verify your setup:

# Capture the exact spectrum, frequencies, and time bins from specgram
Pxx, freqs, time_bins, im = plt.specgram(y, NFFT=256, Fs=sr, noverlap=128)

# Convert Pxx to dB and plot, or reuse the im object's normalization settings
spectrum_db = 10 * np.log10(Pxx)
plt.imshow(spectrum_db, aspect="auto", origin="lower", cmap="viridis")

3. Manual Clipping & Normalization (If You Avoid dB)

If you prefer not to use decibels, clip the spectrum to ignore extreme noise and normalize the remaining values:

raw_spectrum = np.abs(D)
# Clip to the 5th-95th percentile to cut off noise floor and extreme peaks
p_low, p_high = np.percentile(raw_spectrum, [5, 95])
clipped_spectrum = np.clip(raw_spectrum, p_low, p_high)
# Normalize to 0-1 for consistent colormap scaling
normalized_spectrum = (clipped_spectrum - p_low) / (p_high - p_low)

plt.imshow(normalized_spectrum, aspect="auto", origin="lower", cmap="viridis")
plt.show()

4. Use specgram’s Default Colormap

Older versions of Matplotlib used jet for specgram; newer versions use viridis. Make sure your custom plot uses the same colormap to match the visual style.

Final Note #

The big difference is that plt.specgram() is a high-level tool that handles all the messy preprocessing for readable spectral plots. By replicating its internal steps—especially converting to decibels—you’ll get the same clear, well-contrasted visualization as the default specgram output.

内容的提问来源于stack exchange，提问作者Beginner