How Diphone Concatenation Improved Early TTS Systems

Diphone concatenation marked a major breakthrough in early concatenative Text-to-Speech (TTS) synthesis by resolving the severe boundary distortions inherent in basic phoneme-based systems. This article examines how shifting the splice points from phoneme boundaries to phoneme centers allowed synthesizers to preserve natural coarticulation, significantly smoothing acoustic transitions and boosting the intelligibility of synthetic speech.

The Failure of Monophone Concatenation

Early concatenative synthesizers attempted to assemble speech using individual phonemes (monophones) as the basic building blocks. While theoretically simple, this approach produced robotic, jarring speech filled with audible clicks, pops, and unnatural spectral jumps.

The primary cause of this poor quality was coarticulation—the continuous phonetic blending that occurs in natural human speech, where the vocal tract shapes one sound while anticipating the next. In monophone synthesis, cutting and splicing audio at the boundary between two adjacent phonemes meant severing these critical transitional dynamics. Because speech features change most rapidly at the borders between sounds, mismatching the spectral shapes and energy levels at these junctures caused severe acoustic discontinuities.

The Diphone Solution: Shifting the Splice Point

A diphone is an acoustic unit that extends from the center of one phoneme to the center of the next phoneme. Instead of dividing speech at the boundaries between sounds, diphone synthesis shifts the boundaries into the middle of the phones themselves.

Individual Phones:  [  Phone A  ] [  Phone B  ]
                                ^
                     Harsh splice boundary

Diphone Unit:              [ Phone A -> Phone B ]
                                  ^
                        Natural coarticulation preserved

By defining the unit this way, diphone concatenation fundamentally improved acoustic transitions through two primary mechanisms:

1. Preserving Coarticulation Dynamics: The transition between Phone A and Phone B is contained entirely inside the pre-recorded diphone unit. Because this transitional segment was recorded from a continuous human utterance, all the natural formant transitions, spectral trajectories, and consonant-vowel interactions remain intact.
2. Concatenating at Steady States: The splice points are moved to the midpoint of each phoneme, which represents its most acoustically stable, "steady-state" region. Splicing two audio segments together where spectral change is minimal makes acoustic mismatches far less audible to the human ear. Signal processing algorithms, such as Pitch-Synchronous Overlap and Add (PSOLA), could then easily smooth out minor differences in pitch and amplitude across these stable points without degrading the perception of the sound.

Practical Impact on Speech Synthesis

By capturing the transition rather than cutting through it, diphone concatenation created speech that was substantially more intelligible, consistent, and natural than monophone systems. It offered a manageable inventory size—typically requiring between 1,500 and 2,500 units to cover all possible sound pairings in a language like English. This made it computationally viable on early hardware while establishing the foundational principles of concatenative synthesis that paved the way for modern, large-scale unit-selection systems.