diphonebased

Diphone-based synthesis is a form of concatenative speech synthesis that builds speech by concatenating diphones, short audio segments that cover the transition between two adjacent phonemes. A diphone comprises the end of one phoneme and the beginning of the next, capturing coarticulatory transition to produce smoother speech than using isolated phonemes alone.

To synthesize speech, a diphone inventory is created by recording a speaker (or multiple speakers) and labeling units for every ordered pair of phonemes in the target language. For a given text, the system converts the text to a phoneme sequence and selects the corresponding diphones from the inventory. The selected segments are joined with crossfades or overlap-add techniques to minimize discontinuities. Prosody, including duration, stress, pitch, and intonation, can be controlled at the phoneme or diphone level and applied during or after concatenation. Duration models or timing rules adjust the length of each diphone to fit the intended rhythm.

Advantages of diphone-based synthesis include a compact database and relatively natural-sounding transitions for many phrases, making it suitable for devices with limited storage. Limitations include gaps in coverage for less common phoneme sequences, potential artifacts at diphone boundaries if the samples do not match neighboring context, and a more limited expressive range compared to more modern methods.

Historically, diphone-based systems were prominent in early text-to-speech development (1980s–1990s). Although many contemporary systems use unit selection, statistical parametric methods, or neural TTS, diphone-based approaches remain of interest for small-footprint applications and for languages with limited annotated data.