Dynamic Range Compression in TTS Spectrograms

Dynamic range compression is a critical mathematical transformation used when converting raw audio waveforms into log-mel spectrograms for Text-to-Speech (TTS) synthesis. This process balances the vast amplitude disparities inherent in acoustic signals by applying non-linear scaling, typically through logarithmic or power-law operations. By compressing the dynamic range, acoustic feature extractors align synthetic representations with human auditory biology, enhance low-energy phonetic details, and ensure numerical stability during neural network training.

Perceptual Alignment with Human Hearing

Human auditory perception of loudness is non-linear and closely approximates a logarithmic scale, a principle formalized by the Weber-Fechner law. While raw audio pressure variations span several orders of magnitude, the human ear distinguishes relative fractional changes in sound energy rather than absolute linear differences.

Applying dynamic range compression transforms raw linear energy values into a scale that reflects how humans actually perceive speech. Without this step, a TTS model would optimize heavily for loud sound events while treating perceptually significant quiet sounds as negligible background noise.

Preservation of Low-Energy Phonetic Information

Human speech consists of both high-amplitude sounds (such as voiced vowels and fundamental frequencies) and low-amplitude sounds (such as unvoiced fricatives, plosive bursts, and subtle formant transitions). In an uncompressed, linear mel-spectrogram, dominant vowel formants occupy the vast majority of the dynamic range, effectively flattening low-energy components to near-zero values.

Dynamic range compression boosts quieter acoustic details relative to dominant energy peaks. This ensures that essential phonemic cues—such as the subtle high-frequency hiss of an "s" or the release burst of a "t"—remain clearly defined in the time-frequency representation, allowing the TTS acoustic model to learn and synthesize intelligible, natural articulation.

Numerical Stability and Optimization in Neural Networks

From a deep learning perspective, raw power or magnitude spectrogram values create severe optimization challenges:

• Gradient Instability: Large variations in input magnitudes (ranging from \(10^{-7}\) to \(10^{3}\)) lead to exploding or vanishing gradients during backpropagation.
• Loss Function Bias: When computing loss metrics (such as L1 or Mean Squared Error) between predicted and ground-truth spectrograms, high-amplitude frequency bins disproportionately dominate the loss calculation, causing the model to ignore synthesis errors in quieter frequency bands.

Compressing the dynamic range—commonly using a natural logarithm or a base-10 logarithm bounded by a small constant \(\epsilon\) (e.g., \(\log(\max(x, 1e-5))\))—maps values to a well-behaved numerical range, typically between \([-11, 4]\). This normalization creates a smoother loss landscape, accelerates model convergence, and prevents numerical underflow.

Compatibility with Neural Vocoders

Modern neural vocoders (such as HiFi-GAN, WaveGlow, or BigVGAN) are trained to invert log-mel spectrograms back into time-domain waveforms. Dynamic range compression provides these decoders with an information-dense representation where phase-correlated energy distributions are consistently scaled across all frequency bins. This consistency minimizes audible artifacts, phase distortion, and robotic timbre in the final generated speech.