Speech is traditionally modeled as the output of a source-filter system: the glottis (source) produces a series of air pulses, which are then shaped by the vocal tract (filter). A fundamental challenge in speech processing is isolating the source from the filter. While the vocal tract's resonances—formants—shift constantly during speech, the excitation source contains vital information about a speaker's identity, emotion, and physiological health.
Zero-Frequency Filtering (ZFF) has emerged as a robust technique to perform this isolation by exploiting a simple but powerful observation: the impulsive nature of glottal closure creates energy across the entire frequency spectrum, including at 0 Hz.
The Intuition of the Zero Frequency
Traditional methods like inverse filtering attempt to "undo" the vocal tract's effect, which requires an accurate model of the vocal tract itself. However, as Sudarsana Reddy Kadiri et al. note in Breathy to Tense Voice Discrimination using Zero-Time Windowing Cepstral Coefficients (ZTWCCs), the impulsive nature of glottal excitation is observable at all frequencies.
By focusing on the 0 Hz component, researchers can bypass the complex, time-varying resonances of the vocal tract. Because the vocal tract resonances are generally located at much higher frequencies, a filter centered at zero Hz effectively "ignores" the mouth and nose's shaping effects, leaving behind a signal dominated by the timing and strength of the vocal fold vibrations.
The Mechanism of ZFF
The ZFF procedure involves three primary stages to transform raw speech into a clear representation of glottal activity:
- Differentiation: The speech signal is often differenced to reduce the influence of low-frequency components not related to the excitation.
- Cascaded Resonators: The signal is passed through a cascade of ideal digital resonators (integrators) centered at 0 Hz. As described by Eklavya Sarkar et al. in Unsupervised Voice Activity Detection by Modeling Source and System Information using Zero Frequency Filtering, this process highlights the local discontinuities caused by glottal closures.
- Trend Removal: Because ideal integration causes the signal to grow polynomially, a trend removal operation (subtracting the local mean) is applied.
The resulting Zero-Frequency Filtered Signal (ZFFS) oscillates at the fundamental frequency. The positive zero-crossings of this signal correspond to the Glottal Closure Instants (GCIs), also known as epochs. Nagaraj Adiga et al. demonstrate in Zero Frequency Filter Based Analysis of Voice Disorders that these epochs are highly stable even in the presence of noise, as the 0 Hz band is less susceptible to external interference than the spectral regions used by traditional pitch trackers.
Characterizing the Excitation Source
Once the epochs are identified, ZFF allows for the extraction of high-resolution features that describe the "quality" of the voice:
- Strength of Excitation (SoE): The slope of the ZFFS at the zero-crossing indicates how abruptly the vocal folds close. S. R. M. Prasanna et al. used this in Analysis of excitation source information in emotional speech to show that emotions like anger produce significantly stronger excitations than neutral speech.
- Energy of Excitation (EoE): This measures the signal energy in a narrow window around the epoch. Bhanu Teja Nellore et al. utilized these features in Excitation Source and Vocal Tract System Based Acoustic Features for Detection of Nasals in Continuous Speech to distinguish nasal sounds from other voiced segments.
- Instantaneous Pitch: By measuring the time between successive zero-crossings, ZFF provides a cycle-to-cycle estimate of fundamental frequency, which is crucial for analyzing non-stationary signals like laughter. Vinay Kumar Mittal et al. found in Study of changes in glottal vibration characteristics during laughter that laughter involves a reduced closed phase and higher pitch variations compared to modal speech.
Clinical and Practical Applications
The robustness of ZFF makes it particularly valuable for analyzing pathological speech, where traditional models often fail due to high aperiodicity or breathiness. J.C. Vásquez-Correa et al. showed in On Modeling Glottal Source Information for Phonation Assessment in Parkinson’s Disease that ZFF-derived perturbation features are highly effective for assessing the severity of Parkinson’s Disease. Similarly, C M Vikram et al. applied ZFF to detect glottal activity errors in children with cleft lip and palate in Detection of Glottal Activity Errors in Production of Stop Consonants in Children with Cleft Lip and Palate.
Beyond clinical use, ZFF serves as a "super-smoother" for prosodic analysis. Jinfu Ni et al. demonstrated in Using Zero-Frequency Resonator to Extract Multilingual Intonation Structure that the technique can isolate global intonational movements from local micro-prosody, aiding in the development of more natural multilingual text-to-speech synthesis. By focusing on the fundamental physics of the glottis rather than the complex geometry of the vocal tract, ZFF provides a clear window into the engine of human speech.
Key Foundational Work
Analysis of laugh signals for detecting in continuous speech (2009) — contributed 1 concept elements to this direction.
Detection and removal of switching noise in push-to-talk and voice operated exchange communications systems (2007) — contributed 1 concept elements to this direction.