Bark and Mel Scales
- Introduced by
- Eberhard Zwicker, Stanley Smith Stevens
- Year
- 1937
- Type
- Perceptual frequency mapping
- Family
- Psychoacoustics
What does it do?
Bark and Mel scales are perceptual frequency scales that map physical frequency (Hz) to perceived pitch and auditory perception.
Formalized by Zwicker (Bark, 1961) and Stevens (Mel, 1937), these non-linear scales reflect how the human ear processes sound. Bark scale divides hearing into 24 critical bands; Mel scale models pitch perception. Both are essential for audio feature extraction, speech processing, and designing audio systems that align with human hearing.
The intuition behind the method
The ear does not perceive frequency linearly. A 100 Hz shift from 100 Hz to 200 Hz sounds like a large change (octave), but the same shift from 5100 Hz to 5200 Hz is barely noticeable. Bark and Mel scales map frequency to a perceptual space where equal distances represent equal perceived differences. The Bark scale focuses on frequency discrimination (masking patterns); the Mel scale focuses on pitch (melody and tone).
History
Stevens, Volkmann, and Newman introduced the Mel scale in 1937 to quantify pitch perception in equal-appearing intervals (melody). Zwicker developed the Bark scale in 1961, based on critical bandwidth measurements and masking phenomena. Both became foundational to psychoacoustics and audio codec design. Modern refinements (e.g., Gammatone filterbanks, Mel-spectrograms) provide more accurate perceptual models. Machine learning on Mel-spectrograms has become standard in speech and music AI.
When to use it
Use Bark scale for audio codec design, perceptual audio coding, and analyzing masking effects. Use Mel scale for speech processing, music information retrieval, speaker recognition, and machine learning on audio features. Both are applicable to hearing aid design, audio restoration, and perceptual audio quality metrics.
Limitations and considerations
- Bark and Mel scales are empirical approximations; exact critical bandwidth varies with level, frequency context, and individual listeners.
- Non-invertible mappings (forward transformation loses information); precise frequency reconstruction is complex and may introduce artifacts.
- Both scales are average models for normal-hearing listeners; hearing loss, age, and individual differences modulate critical bandwidth and pitch perception.
- Computational cost: converting between Hz and Bark/Mel requires transcendental functions (logarithm, arctangent); lookup tables and approximations are common.
- Limited to monaural (one-ear) perception; binaural hearing (two ears) and spatial audio require additional modeling.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Align analysis with human auditory perception; features derived in Bark/Mel space correlate better with subjective audio quality and listener preference.
- Compress high-frequency information: fewer perceptual 'channels' are needed at high frequencies, reducing computational cost in audio codecs.
- Critical bandwidth concept predicts masking: regions within one Bark are subject to shared masking, simplifying perceptual codec design.
- Well-established and standardized; widely used in commercial audio codecs (MP3, AAC) and speech recognition systems.
- Flexible: Bark and Mel scales can be combined with other techniques (loudness curves, compression) for refined perceptual modeling.
Applications
- Audio codec and streaming: MP3, AAC, OPUS use Bark-scale filterbanks for perceptual quantization; Mel-scale variants used in some modern codecs.
- Automatic speech recognition and speaker verification: Mel-frequency cepstral coefficients (MFCCs) derived from Mel-scale spectrograms are standard features for speech processing.
- Music information retrieval and genre classification: Mel-spectrograms enable music feature extraction and deep learning on audio.
- Hearing aid and cochlear implant design: critical bandwidth (Bark scale) informs channel design and frequency allocation in implants.
- Audio restoration and speech enhancement: perceptual scales guide artifact removal and noise suppression aligned with hearing.
- Audio quality assessment: perceptual metrics (PESQ, PEMO-Q) use Bark and Mel scales to predict subjective audio quality.
How it works
How it works
Acquire an audio signal or frequency-domain representation (FFT spectrum) to be analyzed or displayed.
Compute the power spectrum using standard windowing and FFT techniques. Prepare frequency array in Hz.
Convert linear frequency scale (Hz) to perceptual scales: Bark: z = 13 * arctan(0.76f/1000) + 3.5 * arctan((f/7500)²), Mel: m = 2595 * log10(1 + f/700). These formulae map Hz → perceptual units.
Use perceptual frequency scale to design frequency-selective features (filterbanks, spectrograms) aligned with auditory processing. Compute critical bandwidth at each frequency: BW ≈ 25 + 75(1 + 1.4(f/1000)²)^0.69 in Hz, which in Bark space is approximately constant (1 Bark ≈ critical bandwidth).
Visualize spectra in Bark or Mel scale for perceptually-accurate interpretation. Compare feature importance: in linear Hz, high frequencies dominate visually but may be less perceptually important; in Bark/Mel, emphasis shifts to perceptually salient regions.
Common pitfalls
- Assuming Bark and Mel scales are perceptually identical; they model different aspects (frequency discrimination vs. pitch), and choice depends on application.
- Using a single Bark/Mel scale for all frequencies; critical bandwidth varies with frequency, and some refined models (e.g., Equivalent Rectangular Bandwidth) are more accurate.
- Neglecting level-dependent effects: critical bandwidth and masking scale with sound level; constant critical bandwidth assumptions are approximations.
- Over-compressing high frequencies using Bark scale without perceptual validation; perceived audio quality depends on overall context, not just frequency warping.
- Confusing Bark/Mel conversion with perceptual quality; transforming to Bark scale does not guarantee perceptually optimal audio, only alignment with critical bands.
Frequently asked questions
What is the difference between Bark and Mel scales?
Bark scale models frequency discrimination and critical bandwidth—how the ear groups nearby frequencies into masking bands. Mel scale models pitch perception—how frequency maps to perceived melody. Bark scale is uniform-ish in the perceptual sense (1 Bark ≈ one critical band); Mel scale models subjective pitch (octaves, musical intervals). Use Bark for codec and masking analysis; Mel for pitch and speech processing.
What does one Bark band represent in Hz?
One Bark is not constant in Hz; it represents one critical bandwidth, which varies with frequency. At 100 Hz, 1 Bark ≈ 100 Hz; at 1 kHz, 1 Bark ≈ 130 Hz; at 10 kHz, 1 Bark ≈ 700+ Hz. This means the ear resolves low-frequency details finely but groups high-frequency details coarsely.
How many Bark bands are in human hearing?
Human hearing spans roughly 24 Bark bands, corresponding to critical bandwidth units across the audible spectrum (≈20 Hz to 20 kHz). This is the basis for many audio codecs, which use 24-band filterbanks aligned to Bark spacing.
Why are Mel-spectrograms so popular in speech recognition?
Mel-spectrograms compress high-frequency detail while preserving speech-relevant low-frequency information (pitch, formants). They align with how humans perceive speech, making learned features more robust. Deep neural networks trained on Mel-spectrograms generalize better across speakers and accents.
Can I combine Bark and Mel scales in one analysis?
Yes. Use Bark scale for codec and masking analysis; compute Mel-scale features separately for pitch and melody extraction. They are complementary: Bark for hearing mechanics, Mel for perceptual pitch. Some advanced audio systems use both, layering frequency discrimination (Bark) with pitch perception (Mel).
Related methods
Sources
Zwicker, E. (1961). Subdivision of the audible frequency range into critical bands. Journal of the Acoustical Society of America, 33(2), 248–248.
Open publication referenceCopied.Stevens, S. S., Volkmann, J., & Newman, E. B. (1937). A scale for the measurement of the psychological magnitude pitch. Journal of the Acoustical Society of America, 8(3), 185–190.
Open DOI referenceCopied.Moore, B. C. J. (2012). An Introduction to the Psychology of Hearing (6th ed.). Academic Press.
ISBN: 978-0123914232Copied.