Abnormally low amplitude values in live audio processing

Question

I'm struggling to figure out why my code for live audio feature extraction results in abnormally low amplitude values with little to no variation between frames.

I tried adjusting sampling rate and frame size, which resulted in slightly different amplitude values but still with little to no variation (eg. 0.0032 dB instead of 0.0028 dB).

I tried a different microphone input, and tested both microphones, they seem to function normally.

I tried different approaches to calculate the amplitude (either with numpy RMS, or with librosa library) but that again results in a slightly different amplitude value but still very low and with no variation.

This is my code:

# Parameters
sample_rate = 16000
frame_size = 512

# Initialize PyAudio
p = pyaudio.PyAudio()

# Open a stream for audio input
stream = p.open(format=pyaudio.paInt16,
                channels=1,  # Mono
                rate=sample_rate,
                input=True,
                frames_per_buffer=frame_size)

# Create an empty list to store data rows
data = []
rms_amplitudes_over_time = []
frame_count = 0

try:
    while True:
        # Read audio data from the microphone
        audio_data = np.frombuffer(stream.read(frame_size), dtype=np.int16)

        # Convert audio data to floating-point format
        audio_data = audio_data.astype(np.float32) / 32767.0

        # Calculate F0 (Fundamental Frequency)
        f0 = np.argmax(np.abs(scipy.signal.find_peaks(audio_data)[0]))

        # Calculate RMS Amplitude
        rms_amplitude = np.sqrt(np.mean(librosa.feature.rms(y=audio_data)))
        rms_amplitudes_over_time.append(rms_amplitude)

        # Calculate Shimmer
        peaks, _ = scipy.signal.find_peaks(audio_data)
        amplitude_peaks = np.max(audio_data[peaks]) - np.min(audio_data[peaks])
        shimmer = (amplitude_peaks / len(peaks)) * 100  # Calculate percentage shimmer

        # Calculate Jitter
        periods = np.diff(peaks) / sample_rate
        jitter = np.mean(np.abs(periods - np.mean(periods))) * 1000  # Calculate jitter in milliseconds

        # Extract MFCC features from the audio data
        nfft = 2048  # You can adjust this value as needed
        mfcc_features = mfcc(audio_data, sample_rate, numcep=15, nfft=nfft)

        frame_count += 1
        row = [frame_count, f0, rms_amplitude, shimmer, jitter] + [round(value, 2) for value in mfcc_features[0]]

        # Append the row to the data list
        data.append(row)

        # Print the current frame's data
        current_frame_data = [row]
        current_df = pd.DataFrame(current_frame_data, columns=["Frame", "F0 (Hz)", "RMS Amplitude", "Shimmer (%)", "Jitter (ms)"] + [f"MFCC {i}" for i in range(15)])
        print(current_df)

        # Sleep briefly to control the display rate (adjust as needed)
        time.sleep(1.0)

except KeyboardInterrupt:
    # Press Ctrl+C to stop the script
    pass

# Create a DataFrame from the collected data
columns = ["Frame", "F0 (Hz)", "RMS Amplitude (dB)", "Shimmer (%)", "Jitter (ms)"] + [f"MFCC {i}" for i in range(15)]
df = pd.DataFrame(data, columns=columns)

# Print the final DataFrame
print(df)

# Visualize RMS amplitude over time
plt.plot(rms_amplitudes_over_time)
plt.xlabel("Frame")
plt.ylabel("RMS Amplitude")
plt.title("RMS Amplitude Over Time")
plt.show()

# Cleanup and close the audio input stream
stream.stop_stream()
stream.close()
p.terminate()