Preparing a Text-to-Speech Dataset

A TTS training example connects an audio segment to the text intended to produce it. The pairing must be accurate: if the recording contains a different word, missing phrase, or long silence, the model receives conflicting supervision. Before training, listen to samples, compare waveforms with boundaries, and spot-check transcripts rather than trusting an automated manifest.

Recording conditions should be stable enough that the model can learn the intended voice rather than changing microphones or rooms. Use a consistent sample rate, channel configuration, distance, and gain. Avoid clipping, abrupt noise, and overlapping speakers. Overprocessing can remove natural prosody or introduce artifacts, so denoising and loudness normalization should be conservative and documented. Preserve original recordings and processing provenance.

Segmentation should produce clips with natural linguistic boundaries and enough context for the model. Very short fragments may omit coarticulation, while long clips increase alignment and memory challenges. Avoid clipping initial consonants or final phonemes; include meaningful breaths only when the annotation convention and model support them, and preserve punctuation cues where expected. If forced alignment is used, inspect uncertain segments and adapt boundaries to the model's expected format.

Text normalization conventions must be consistent. Decide how to represent numbers, abbreviations, punctuation, disfluencies, and non-speech vocalizations. The written transcript need not mimic orthography identically across all languages, but it must match the model's tokenization or phonemization assumptions. Pronunciation dictionaries or phoneme labels may be needed for uncommon names.

Measure speaker and phonetic coverage, not just total hours. A dataset dominated by repetitive phrases may leave rare sounds unrepresented. Choose a split unit that matches the generalization question: hold out speakers when testing unseen-speaker performance, or hold out sessions and text for same-speaker adaptation. Keep each augmented copy in its source example’s partition, and keep evaluation text separate from training. A manifest format can help training code locate files, but schema compliance does not guarantee annotation correctness or voice consent.