Speech Enhancement Evaluation Metrics
1. Intrusive Metrics¶
These metrics require the presence of clean samples (ground truth) against which the generated speech is evaluated.
Speech Quality¶
- PESQ (Perceptual Evaluation of Speech Quality)
A legacy metric comparing a reference against a degraded signal by modeling the human auditory system.
- Range: -0.5 to 4.5.
- Context: Originally for narrow-band (8kHz) and wide-band (16kHz) telecommunications. Succeeded by POLQA.
- POLQA (Perceptual Objective Listening Quality Analysis)
The successor to PESQ, assessing super-wideband (32kHz) or fullband (48kHz) signals.
- Advantage: Less susceptible to time-warping or slight phase misalignments induced by modern buffering and AI codecs.
- ViSQOL A spectro-temporal metric introduced by Google as a light-weight alternative to POLQA, addressing similar issues found in PESQ.
Noise Suppression (Composite Scores)¶
These scores evaluate noise and signal quality separately, often mixing PESQ with SNR and Log-Likelihood metrics. * CSIG: Signal quality score. * CBAK: Noise suppression score. * COVL: Overall weighted average (Composite Overall). * > [!NOTE] > These metrics share similar limitations and issues with PESQ.
Speech Intelligibility¶
- STOI (Short-Time Objective Intelligibility): Measures the correlation of temporal envelopes in short-time segments, crucial for word recognition.
- ESTOI (Extended STOI): Better handles background noise like "babble," making it superior for complex noise scenarios.
- MBSTOI (Modified-Binaural-STOI): Considers the human ability to leverage binaural cues, suited for multi-channel binaural speech enhancement.
Other Spectral Metrics¶
- MCD (Mel Cepstral Distortion) Measures distance between mel-cepstral coefficients; standard for assessing timbre and speaker identity retention.
- LSD (Log-Spectral Distance) Measures distance between log-spectra. Helps detect high-frequency oversmoothing that PESQ often misses.
2. Non-Intrusive Metrics¶
These metrics do not require a ground-truth reference for comparison.
Speech Quality¶
- DNSMOS (Deep Noise Suppression Mean Opinion Score) A CNN-based model predicting human ratings. Provides scores for SIG (Signal), BAK (Background), and OVRL (Overall). Preferred over PESQ/POLQA for modern AI models.
- NISQA (Non-Intrusive Speech Quality Assessment) A deep learning framework predicting MOS for overall quality and specific dimensions (noisiness, coloration, discontinuity, and loudness).
- UTMOS / SSL-MOS Newer metrics based on Self-Supervised Learning (SSL) backbones (like wav2vec 2.0) fine-tuned on human subjective ratings.
Speech Intelligibility¶
- SRMR (Speech-to-Reverberation Modulation Energy Ratio) Measures intelligibility specifically in reverberation scenarios.
3. Audiological Metrics¶
Used for medical assessment of speech comprehension in individuals, particularly hearing-impaired patients.
Core Assessments¶
- SRT (Speech Reception Threshold) The dB level at which 50% of speech material is recognized correctly.
- WRS (Word Recognition Score) The percentage of words correctly identified at a comfortable volume (~30-40 dB above SRT).
- SIN (Speech-in-Noise) The SNR at which a listener understands speech in a noisy environment. Often utilizes multi-speaker lab setups or VR environments.
Predictive Models for Hearing Loss¶
- HASPI (Hearing Aid Speech Perception Index) Models the human auditory periphery (including hair cell damage) to predict intelligibility for people with hearing loss.
- HASQI (Hearing Aid Speech Quality Index) Predicts how "natural" or "pleasant" audio sounds to someone with a specific hearing profile.