Synthetic speech makes it possible to build a training corpus with consistent, paired voices without recording real speakers. The key trade-off is that this consistency comes with the pronunciation habits and errors of the model that generated the audio.
What Makes It Useful
- Each sentence is read in both a female and a male voice, making the corpus useful for studying or training two-voice TTS without speaker recordings.
- The 17,371 clips span 19.1 hours, with text sources including LibriTTS-R, Common Voice, project-written templates, and LLM-written sentences.
- Phoneme, stress, and duration annotations support training pipelines that need more than raw audio-text pairs.
- Sentence-level train, validation, and test splits keep both voice readings together, reducing the chance of near-duplicate text leaking across splits.
Who It Suits—and the Trade-Off
A good fit for prototyping English TTS systems, testing data pipelines, or exploring phoneme-aligned synthetic speech under CC BY 4.0. Look elsewhere for natural recorded speech, many speakers or accents, conversational spontaneity, or broad coverage of digits and symbols: the corpus has just two synthetic voices, one US accent, and read speech, and its automatic checks do not establish naturalness.