voicebank

A voicebank is a collection of recorded voice samples and the accompanying metadata used to synthesize speech or singing. It represents the voice data of a particular speaker or singer and serves as the input material for a synthesis engine, which combines the samples to produce new utterances or melodies.

A typical voicebank includes audio recordings, a pronunciation or phoneme inventory, alignment information that links sections of audio to phonemes or notes, and descriptive metadata about the speaker such as language, dialect, gender, and age. For singing voicebanks, additional data often accompany the samples, including note annotations, lyrics alignment, and pitch or vibrato curves that guide the musical output. The data are usually organized to support efficient retrieval by phoneme, pitch, or timing.

Creation and quality control are central to voicebank development. Voice actors record scripts provided by the project, sometimes under controlled acoustics. Recordings are processed to reduce noise, normalize levels, and ensure consistent timing. Audio is labeled and aligned with phonetic or musical targets, and the resulting dataset is cataloged with licensing terms and usage rights. Ethical and legal considerations include obtaining consent and ensuring proper attribution and licensing for redistribution and reuse.

Voicebanks are used in speech synthesis and singing synthesis, including technologies that concatenate samples or model voice characteristics to generate new audio. They may be distributed as standalone datasets or as part of a larger synthesis system, and formats vary by platform, with common elements being audio files and accompanying text or annotation files describing phoneme, timing, and pitch information.