Audio Quality Metrics: A Comprehensive Reference

Description: Measures the active speech level excluding pauses and silent segments.

How it works: Applies voice activity detection (VAD) to isolate speech regions, then calculates RMS level of active segments.

Formula: \(\text{ASL} = 10 \log_{10} \left( \frac{1}{N} \sum_{i=1}^{N} x_i^2 \right)\) where \(x_i\) are active speech samples.

Libraries:

• Python: pydub, librosa
• MATLAB: Audio Toolbox

References:

• ITU-T P.56: Objective measurement of active speech level

Description: Perceptually weighted loudness measurement for broadcast audio.

How it works: Applies K-weighting filter to approximate human loudness perception, integrates over time.

Libraries:

• Python: pyloudnorm
• C++: libebur128

Datasets:

• EBU Loudness Test Set

References:

• ITU-R BS.1770-4: Algorithms to measure audio programme loudness

Description: Ratio of signal power to noise power, expressed in dB.

How it works: \(\text{SNR} = 10 \log_{10} \left( \frac{P_{\text{signal}}}{P_{\text{noise}}} \right)\)

Limitations: Does not correlate well with perceptual quality.

Libraries:

• Python: scipy.signal, numpy
• MATLAB: Built-in snr() function

References:

• IEEE Standard for Audio Quality Measurement

Description: ITU standard for objective audio quality measurement, designed for codec evaluation.

How it works: Psychoacoustic model comparing reference and degraded signals across frequency bands.

Libraries:

• C: GstPEAQ (GStreamer plugin)
• MATLAB: Third-party implementations available

Datasets:

• MUSHRA Test Corpus

References:

• ITU-R BS.1387-1: Method for objective measurements of perceived audio quality

Description: ITU standard for predicting speech quality in telecom networks.

How it works: Time-aligned comparison of reference and degraded signals through perceptual model. Output: MOS-LQO (0.5 to 4.5 scale).

Libraries:

• Python: pesq (via pip)
• C: Official ITU implementation

Datasets:

• NOIZEUS Speech Corpus
• TIMIT Acoustic-Phonetic Corpus

References:

• ITU-T P.862: Perceptual evaluation of speech quality (PESQ)
• Rix, A.W., et al. (2001). “Perceptual evaluation of speech quality (PESQ) – a new method for speech quality assessment of telephone networks and codecs”

Description: Successor to PESQ, supporting wideband and super-wideband speech.

How it works: Advanced perceptual model with improved handling of time warping and codec artifacts.

Libraries:

• Commercial: POLQA by OPTICOM
• Python: Limited open-source implementations

References:

• ITU-T P.863: Perceptual objective listening quality assessment

Description: Predicts speech intelligibility in noisy conditions.

How it works: Correlates time-frequency representations of clean and processed speech.

Libraries:

• Python: pystoi
• MATLAB: STOI Toolbox

Datasets:

• LibriSpeech
• DNS Challenge Dataset

References:

• Taal, C.H., et al. (2011). “An Algorithm for Intelligibility Prediction of Time–Frequency Weighted Noisy Speech”

Description: Measures perceptual contrast enhancement in processed speech spectrograms.

How it works: Computes contrast ratio in time-frequency domain weighted by auditory masking.

Libraries:

• Custom implementations (research-specific)

References:

• Healy, E.W., et al. (2013). “An algorithm to increase speech intelligibility for hearing-impaired listeners”

Description: Deep learning-based predictor of subjective MOS for noise suppression systems.

How it works: Neural network trained on large-scale listening tests to predict MOS directly from audio.

Libraries:

• Python: DNSMOS by Microsoft

Datasets:

• Microsoft DNS Challenge Dataset

References:

• Reddy, C.K.A., et al. (2021). “DNSMOS: A Non-Intrusive Perceptual Objective Speech Quality metric”

Description: ANSI standard for predicting speech intelligibility based on audibility.

How it works: Weights audible speech bands according to their importance for intelligibility.

Libraries:

• MATLAB: SII Toolbox

References:

• ANSI S3.5-1997: Methods for Calculation of the Speech Intelligibility Index

Description: Extension of STOI handling non-linear processing like spectral subtraction.

How it works: Applies intermediate intelligibility measure to improve correlation with subjective scores.

Libraries:

• Python: pystoi (includes ESTOI)
• MATLAB: ESTOI Toolbox

References:

• Jensen, J., & Taal, C.H. (2016). “An Algorithm for Predicting the Intelligibility of Speech Masked by Modulated Noise Maskers”

Description: Non-intrusive metric estimating intelligibility degradation due to reverberation.

How it works: Analyzes modulation spectrum energy ratio across frequency bands.

Libraries:

• MATLAB: SRMR Toolbox

Datasets:

• ACE Challenge Corpus

References:

• Falk, T.H., et al. (2010). “A non-intrusive quality and intelligibility measure of reverberant and dereverberated speech”

Description: Time for sound to decay by 60 dB after source stops.

How it works: Measures decay slope of impulse response in frequency bands.

Libraries:

• Python: pyroomacoustics
• MATLAB: ITA-Toolbox

Datasets:

• ACE Corpus
• BUT Speech@FIT Reverb Database

References:

• ISO 3382-1: Measurement of room acoustic parameters

Description: Ratio of early to late arriving sound energy.

How it works: \(C_{50} = 10 \log_{10} \left( \frac{\int_0^{50ms} p^2(t) dt}{\int_{50ms}^{\infty} p^2(t) dt} \right)\)

Applications: Speech clarity (C50), music clarity (C80).

Libraries:

• Python: pyroomacoustics

References:

• ISO 3382-1: Measurement of room acoustic parameters

Description: Predicts speech intelligibility for hearing-impaired listeners with and without hearing aids.

How it works: Models auditory processing including hearing loss and amplification.

Libraries:

• MATLAB: HASPI Toolbox

References:

• Kates, J.M., & Arehart, K.H. (2014). “The Hearing-Aid Speech Perception Index (HASPI)”

Description: Measures separation quality as ratio of target signal to artifacts.

How it works: Decomposes error into target distortion, interference, and noise.

Libraries:

• Python: mir_eval.separation

Datasets:

• MUSDB18
• WSJ0-2mix

References:

• Vincent, E., et al. (2006). “Performance measurement in blind audio source separation”

Description: Scale-invariant version of SDR, more robust to amplitude differences.

How it works: Projects estimated signal onto reference, computes distortion ratio.

Libraries:

• Python: torch_mir_eval or custom implementation

References:

• Le Roux, J., et al. (2019). “SDR – half-baked or well done?”

Description: See “Overall Audio Quality” section above.

Description: Perceptual quality metric for speech and audio, supporting music mode.

How it works: Spectrogram-based similarity using neurogram representation.

Libraries:

• C++/Python: ViSQOL by Google

Datasets:

• Custom music test sets (check ViSQOL repo)

References:

• Hines, A., et al. (2015). “ViSQOL: an objective speech quality model”

Description: Euclidean distance between log-magnitude spectra.

How it works: \(\text{LSD} = \sqrt{ \frac{1}{K} \sum_{k=1}^{K} \left( 10 \log_{10} |X_k| - 10 \log_{10} |\hat{X}_k| \right)^2 }\)

Libraries:

• Python: Custom with librosa or numpy

References:

• Gray, A., & Markel, J. (1976). “Distance measures for speech processing”

Description: Distance between mel-frequency cepstral coefficients (MFCCs).

How it works: \(\text{MCD} = \frac{10}{\ln 10} \sqrt{2 \sum_{k=1}^{K} (c_k - \hat{c}_k)^2}\)

Applications: Voice conversion, TTS evaluation.

Libraries:

• Python: librosa, scipy

References:

• Kubichek, R. (1993). “Mel-cepstral distance measure for objective speech quality assessment”

Description: Percentage of word errors (substitutions, deletions, insertions) in ASR output.

How it works: \(\text{WER} = \frac{S + D + I}{N} \times 100\%\) where S=substitutions, D=deletions, I=insertions, N=total words.

Libraries:

• Python: jiwer

Datasets:

• LibriSpeech
• Common Voice

References:

• NIST Speech Recognition Scoring Toolkit

Description: Contextual embedding similarity between reference and hypothesis transcriptions.

How it works: Computes cosine similarity of BERT embeddings token-by-token.

Libraries:

• Python: bert-score

References:

• Zhang, T., et al. (2020). “BERTScore: Evaluating Text Generation with BERT”

Description: Predicts speech quality (not just intelligibility) for hearing aid users.

How it works: Models auditory processing with hearing loss, computes quality along multiple dimensions.

Libraries:

• MATLAB: HASQI Toolbox

References:

• Kates, J.M., & Arehart, K.H. (2010). “The Hearing-Aid Speech Quality Index (HASQI)”

Description: Improvement in intelligibility from binaural vs. monaural listening.

How it works: Compares predicted intelligibility under binaural and monaural conditions.

Applications: Bilateral hearing aid fitting, spatial audio benefits.

References:

• Culling, J.F., et al. (2004). “The role of head-induced interaural time and level differences”

Description: Measures temporal variability in soundscapes, correlates with biodiversity.

How it works: Computes intensity differences across adjacent time frames in frequency bands.

Libraries:

• R: soundecology
• Python: scikit-maad

Datasets:

• Xeno-canto (bird recordings)
• AudioSet

References:

• Pieretti, N., et al. (2011). “A new methodology to infer the singing activity of an avian community”

Description: Ratio of biophony (1-2 kHz) to anthrophony (1-2 kHz and 2-11 kHz).

How it works: \(\text{NDSI} = \frac{\text{Biophony} - \text{Anthrophony}}{\text{Biophony} + \text{Anthrophony}}\)

Libraries:

• R: soundecology
• Python: scikit-maad

References:

• Kasten, E.P., et al. (2012). “The remote environmental assessment laboratory’s acoustic library”

Description: Area under the spectrum curve within frequency range of biological sounds.

How it works: Integrates spectral energy in 2-8 kHz range (typical for birds/insects).

Libraries:

• R: soundecology

References:

• Boelman, N.T., et al. (2007). “Multi-trophic invasion resistance in Hawaii”

Description: Subjective assessment of soundscape quality in urban environments.

How it works: Perceptual attributes evaluated via listening tests (pleasantness, eventfulness, etc.).

Standards:

• ISO 12913-3: Data analysis and reporting

Datasets:

• SATP Database (Soundscapes & Psychophysiology)

References:

• Aletta, F., et al. (2016). “Soundscape descriptors and a conceptual framework for developing predictive soundscape models”