dB SPL vs dBFS: Understanding Audio Levels for Engineers and Clinicians

Two recordings can both show -18 dBFS on a meter and yet represent completely different physical sound levels — one recorded with a studio condenser at arm's length, the other with a clinical-grade microphone 30 cm from a patient's mouth. Without understanding the distinction between dB SPL and dBFS, you cannot compare them, calibrate them, or trust any absolute measurement derived from them.

This article cuts through the confusion. We define both scales precisely, derive the conversion formula, and explain the practical consequences for audio engineers, ML practitioners, and clinical researchers working with voice data.

1. What Is dB SPL?

Sound Pressure Level (SPL) is the physical measure of the pressure fluctuations a sound wave creates in air. It is defined as:

dB SPL = 20 × log10(p / p_ref)

where:
  p      = measured RMS sound pressure (in Pascals)
  p_ref  = 20 µPa  (20 × 10⁻⁶ Pa — the threshold of human hearing)

The reference 20 µPa is not arbitrary — it corresponds to the quietest sound a young, healthy human ear can detect at 1 kHz. This makes 0 dB SPL the absolute threshold of human hearing, not silence.

dB SPL is an absolute physical scale. A measurement of 94 dB SPL describes a real pressure in the world, independent of any microphone or recording system. Here are common reference points:

2. What Is dBFS?

dBFS (decibels relative to Full Scale) is the digital equivalent — a relative scale that describes the amplitude of a digital audio signal with respect to the maximum amplitude a digital system can represent without clipping:

dBFS = 20 × log10(A / A_max)

where:
  A      = measured amplitude (peak or RMS)
  A_max  = maximum representable digital amplitude (full scale)

Because A is always less than or equal to A_max, dBFS values are always negative or zero. The moment a signal hits 0 dBFS, it has reached the ceiling of the digital system — any louder and it clips.

This scale is used everywhere digital audio is processed: DAWs (Pro Tools, Logic, Ableton), audio APIs (PyDub, librosa, soundfile), WAV files, and mobile recording apps all report levels in dBFS. When Python's librosa computes RMS energy from a float32 audio array normalised to [-1, 1], the resulting value in dB is a dBFS value — not a physical sound level.

3. The Key Difference: Absolute vs. Relative

The confusion between dB SPL and dBFS is understandable — both use decibels, both involve logarithms of amplitude ratios, and both can describe audio. But they answer fundamentally different questions:

A practical analogy: dB SPL is like measuring temperature in Kelvin — it references a true physical zero. dBFS is like measuring temperature relative to the highest temperature your thermometer can read. Without knowing where "full scale" sits on the Kelvin axis, the reading is meaningless in absolute terms.

4. Converting Between dB SPL and dBFS

The bridge between the two scales is a single constant called the calibration offset. It represents the dB SPL that corresponds to 0 dBFS in your recording system — in other words, how loud a sound would have to be in the physical world to drive your microphone/ADC chain to full scale:

dB_SPL = dBFS + calibration_offset

Equivalently:
  calibration_offset = dB_SPL_known − dBFS_measured

Example:
  A pistonphone outputs 94 dB SPL.
  You record it and measure -28.4 dBFS (RMS).
  calibration_offset = 94 − (−28.4) = 122.4 dB

  A subsequent recording reads -18.0 dBFS (RMS).
  dB_SPL = -18.0 + 122.4 = 104.4 dB SPL

Python: Calibrated SPL from a WAV File

import numpy as np
import soundfile as sf

def rms_dbfs(signal: np.ndarray) -> float:
    """Compute RMS level in dBFS (full scale = 1.0)."""
    rms = np.sqrt(np.mean(signal ** 2))
    if rms == 0:
        return -np.inf
    return 20 * np.log10(rms)

def dbfs_to_spl(dbfs: float, calibration_offset: float) -> float:
    """Convert dBFS to dB SPL using a known calibration offset.

    calibration_offset = dB_SPL_reference - dBFS_measured_at_reference
    (typically 94 dB SPL pistonphone measured at the recording chain)
    """
    return dbfs + calibration_offset

# --- Usage ---
audio, sr = sf.read("voice_sample.wav")

# Compute level from the full recording
level_dbfs = rms_dbfs(audio)
print(f"Recording level: {level_dbfs:.1f} dBFS")

# Convert to physical SPL (requires prior calibration)
CALIBRATION_OFFSET = 122.4  # dB, determined from pistonphone session
level_spl = dbfs_to_spl(level_dbfs, CALIBRATION_OFFSET)
print(f"Estimated SPL:   {level_spl:.1f} dB SPL")

Note that soundfile reads WAV files as float32 in the range [-1, 1], so the RMS of a full-scale sine wave is 1/√2 ≈ 0.707, giving -3.01 dBFS — consistent with the dBFS definition.

5. Why It Matters for Clinical Audio

Clinical audio applications — voice pathology assessment, speech therapy, acoustic biomarker research — have stricter requirements than most consumer audio use cases. The reason: clinical decisions depend on the accuracy of computed metrics, and those metrics are highly sensitive to recording level.

Acoustic Biomarkers: AVQI, Jitter, Shimmer, HNR

Metrics like the Acoustic Voice Quality Index (AVQI), jitter (cycle-to-cycle frequency variation), shimmer (cycle-to-cycle amplitude variation), and Harmonics-to-Noise Ratio (HNR) are extracted from the raw waveform of a sustained phonation. Their validity depends on:

• No clipping: A signal exceeding 0 dBFS introduces nonlinear distortion. The resulting waveform is no longer a faithful reproduction of the patient's voice — shimmer and jitter values computed from it are meaningless.
• Adequate SNR: A recording level below -40 dBFS (peak) typically means the signal is buried in the noise floor. HNR measurements will reflect the microphone noise rather than the patient's voice quality.
• Calibrated absolute level: Some normative AVQI ranges are established at specific vocal effort levels (e.g., comfortable loudness at 30 cm). Comparing a patient against normative data requires that your recording represents the same physical SPL, not just the same dBFS.

Deep Learning Models: Wav2Vec2 and Input Normalisation

Models like Wav2Vec2 and HuBERT were pre-trained on audio normalised to a specific amplitude range — typically float32 in [-1, 1] with RMS-based loudness normalisation. When you pass a poorly levelled recording (very quiet, or at an unusual gain stage), the model's learned representations may not generalise correctly because the activation statistics at the input convolutional layers differ from the training distribution.

Best practice for inference with these models: normalise your audio to a consistent RMS level (e.g., -23 dBFS per the EBU R128 standard) before passing it to the model, then apply your calibration offset separately if you need absolute SPL. Never conflate the normalisation step with calibration — normalisation is a preprocessing choice; calibration maps your system to the physical world.

6. Practical Rules of Thumb

Frequently Asked Questions