Chest to Falsetto Transitions in Neural Singing TTS

Neural singing Text-to-Speech (TTS) models have advanced significantly, yet achieving a seamless transition between chest voice and falsetto remains a major technical hurdle. This article examines the core acoustic, algorithmic, and data-driven challenges that neural architectures face when synthesizing vocal register shifts, highlighting why bridging the gap between modal voice and falsetto often leads to audible artifacts, pitch instability, or unnatural timbre changes.

Physiological and Acoustic Discontinuity

The transition between chest voice (modal register, or Mechanism 1) and falsetto (loft register, or Mechanism 2) involves a fundamental change in vocal fold biomechanics. In chest voice, the vocal folds are thick and vibrate across their entire depth, producing rich harmonic energy and a low open quotient. In falsetto, the folds elongate, thin out, and vibrate primarily along the edges, resulting in a higher open quotient, steeper spectral tilt, and increased aspiration noise. Neural acoustic models struggle to continuously interpolate these discrete physical states, often generating an unnatural "break" or a sudden collapse in acoustic density rather than a gradual shift.

Feature Space and Vocoder Limitations

Most modern singing TTS pipelines rely on intermediate representations like mel-spectrograms paired with neural vocoders (e.g., HiFi-GAN or diffusion-based vocoders). Mel-spectrograms compress frequency bins, often smoothing out the precise harmonic-to-noise ratio (HNR) shifts critical during the passaggio (the transition zone). When a model attempts to blend registers, vocoders frequently generate phase anomalies, robotic buzziness, or muffled high frequencies because the predicted spectrogram represents an impossible physical average rather than a coherent glottal waveform.

Entanglement of Pitch and Register

In human singing, pitch (fundamental frequency, or \(F_0\)) and vocal register are related but decoupled: a singer can produce the same pitch in chest voice, mixed voice, or falsetto. Neural singing models, however, tend to strongly entangle pitch and timbre due to latent correlation. If the training data predominantly contains chest voice at lower registers and falsetto at higher registers, the model struggles when directed to sustain chest voice higher or enter falsetto lower, leading to predictive confusion and unstable \(F_0\) contours around the register boundaries.

Scarcity of Continuous Passaggio Data

High-quality singing datasets rarely contain sufficient examples of smoothly executed vocal transitions. Professional singers undergo years of training to minimize the audible break between registers using mixed voice (voix mixte). In typical corpora, vocalists either switch abruptly between registers or stick to one dominant mode for a given phrase. Without extensive, multi-pitch, multi-register vocal exercises captured across dynamic ranges, the neural network lacks the ground truth necessary to learn continuous mapping across the transition zone.

Inadequate Input Representation and Control

Standard singing TTS score inputs (lyrics, phonemes, musical note pitches, and durations) lack explicit instructions regarding vocal fold mechanics. Without dedicated conditioning labels—such as explicit register flags, glottal closure parameters, or continuous timbre embeddings—the acoustic model faces an ill-posed problem. Given identical phonetic and pitch inputs at the register boundary, the model is forced to resolve a multi-modal distribution, which frequently manifests as unstable, flickering transitions in the synthesized audio.