Silence-to-Speech Ratio in VAD Systems

• Silence-to-Speech Ratio (SSR) is defined as the ratio of non-speech (silence) frames to speech frames, serving as a key metric in VAD performance evaluation.
• Systematic SSR manipulation in LibriVAD through NonConcat and Concat regimes demonstrates its impact on enhancing model generalization, particularly under low SNR and out-of-distribution conditions.
• Precise SSR control via forced-alignment based silence extraction and calibrated silence insertion improves class balance and robustness, offering actionable insights for real-world VAD deployments.

The silence-to-speech ratio (SSR) is a central corpus property and experimental parameter in the design and benchmarking of Voice Activity Detection (VAD) systems. SSR quantifies the proportion of non-speech (pure silence) frames relative to speech-active frames in a given audio dataset. Contemporary work, especially in the construction and analysis of large-scale VAD corpora such as LibriVAD, demonstrates that SSR is a critical factor influencing model generalization, class balance, and robustness under adverse and out-of-distribution (OOD) conditions (Stylianou et al., 19 Dec 2025).

1. Formal Definition of Silence-to-Speech Ratio (SSR)

SSR is formally defined as the ratio of the total duration of silence frames ($T_{\text{silence}}$) to the total duration of speech frames ($T_{\text{speech}}$):

$$\mathrm{SSR} = \frac{T_{\text{silence}}}{T_{\text{speech}}}$$

For instance, an SSR of 0.25 indicates that for each second of speech, there is 0.25 seconds of silence; equivalently, 20% of the dataset contains silence. Some reports express SSR as a fraction of total frames, but the principal definition remains $T_{\text{silence}} / T_{\text{speech}}$. This formalization is compatible with the binary framing used in VAD tasks, where data are partitioned into speech-active and non-speech (silence) classes (Stylianou et al., 19 Dec 2025).

2. Systematic Control of SSR in LibriVAD

LibriVAD exemplifies systematic SSR manipulation via two "silence-regimes"—NonConcat and Concat—each constructed at three corpus scales (small, medium, large):

• LibriVAD-NonConcat: Uses unmodified LibriSpeech subsets. Forced-alignment indicates roughly 17.6% of frames are non-speech, corresponding to SSR $\approx 0.176/(1-0.176) \approx 0.21$.
• LibriVAD-Concat: Forms consecutive utterance pairs, concatenated and injected with silence such that inserted non-speech equals 25% of combined speech duration. Pure silence segments are mined via forced-alignment from train-clean-100 to form a "silence reservoir." After injection, SSR is increased to approximately 0.34, with ~34% of the dataset comprising non-speech frames.

Both silence regimes are further modified by mixing with nine diverse noise types and sweeping six SNR levels (–5, 0, 5, 10, 15, 20 dB), yielding variant datasets at three total sizes (≈15 GB, ≈150 GB, ≈1.5 TB) to facilitate controlled SSR studies (Stylianou et al., 19 Dec 2025).

Variant	Non-Speech (%)	SSR	Silence Manipulation
NonConcat	17.6%	≈0.21	None
Concat	34%	≈0.34	Silence injection

3. SSR Effects on VAD Performance and Generalization

Extensive experiments involving three model families—raw-CLDNN (waveform), bDNN (MFCC/GFCC), and Vision Transformer (ViT, MFCC/GFCC)—quantify the performance impact of SSR under matched and OOD settings. Performance metrics include AUC, EER, and MinDCF.

• Small-scale training ($\sim$55 h train, 3 h test): ViT+MFCC is optimal. On NonConcat (SSR≈0.18), average AUC is 0.9574; on Concat (SSR≈0.34), AUC increases to 0.9710. Across all model/feature pairs, Concat outperforms NonConcat by 1–2 absolute AUC points, with maximal gains at low SNR (–5 dB) and on unseen noise.
• Scaling training data: Transitioning from small to medium corpora further increases AUC, e.g., ViT+MFCC Concat, 0.9710→0.9761. However, further increase to the largest corpus produces a plateau or minor in-domain AUC drop, but delivers best generalization in cross-domain and real-world OOD (VOiCES) evaluations.
• VOiCES OOD evaluation: Example AUCs—NonConcat small=0.9225, Concat small=0.9356; with large data, NonConcat=0.9560, Concat=0.9666. The higher-SSR Concat corpus provides a systematic generalization advantage in every OOD scenario assessed (Stylianou et al., 19 Dec 2025).

4. Methodology for Precise SSR Construction

Precise SSR control requires frame-accurate silence identification and flexible corpus construction:

• Forced-alignment–based silence extraction is used to obtain pure non-speech frames for injection. This technique enables deterministic manipulation of SSR without manual re-labeling.
• Deterministic silence insertion: In Concat, silence is added between paired utterances by selecting calibrated segments from the silence reservoir. This supports controlled SSR scaling to target ratios.
• Noise and SNR regime diversity: Once speech–silence regimes are specified, mixtures are created with a fixed grid of noise types and SNR levels, ensuring each SSR configuration is evaluated under a broad range of real and simulated adverse conditions (Stylianou et al., 19 Dec 2025).

5. Practical Recommendations and Corpus Design Implications

Key insights from controlling and analyzing SSR in LibriVAD include:

• Class balance optimization: Standard LibriSpeech (~18% non-speech) yields class imbalance. Raising SSR to ~34% via concatenation and silence injection reduces speech false positives by exposing models to diverse silence patterns.
• Optimal SSR range: An SSR of 25–40% achieves a balance between excessive silence and insufficient non-speech, supporting model robustness across conditions. VAD corpus designers are advised to target this range for practical deployments.
• Alignment with deployment scenarios: Aligning SSR in training data with deployment expectations—especially for low SNR/OOD settings—improves generalization. Training sets should match or slightly exceed anticipated non-speech fractions.
• Combination with other robustness strategies: SSR balancing should be integrated with dataset size scaling and noise diversity for maximal effect (Stylianou et al., 19 Dec 2025).

6. Significance and Future Directions

The systematic manipulation and benchmarking of SSR in LibriVAD establishes a new standard for VAD corpus design and evaluation. The central finding is that SSR balancing, combined with large and diverse datasets, is critical for advancing VAD generalization, particularly in challenging acoustic and domain-shift scenarios. By publishing both NonConcat and Concat variants, LibriVAD enables the community to further explore SSR optimization and refine models capable of robust speech–silence discrimination "in the wild" (Stylianou et al., 19 Dec 2025). A plausible implication is the extension of SSR balancing concepts to other binary detection tasks (e.g., music/speech, environmental event detection), where class balance and temporal statistics may similarly affect real-world generalization.