Clean Speech Corpus for Text-to-Speech Synthesis

Cv Tts Clean is a speech dataset for text-to-speech applications, created by neongeckocom and uploaded to Hugging Face in September 2022. The dataset name suggests it contains clean audio recordings, likely paired with corresponding text transcripts. Its BSD 3-Clause license and US region tag indicate permissible use and a primary geographic source.

Use Cases

• Training a TTS model to generate speech waveforms from text transcripts.
• Fine-tuning a vocoder on clean audio features like mel-spectrograms or linguistic units.
• Benchmarking speech synthesis quality using metrics like MOS (Mean Opinion Score) on clean, studio-recorded audio.
• Developing voice cloning systems using speaker embeddings extracted from consistent audio recordings.

Strengths

• Dataset is hosted on Hugging Face, a major platform for ML datasets, ensuring structured access.
• The BSD 3-Clause license provides clear, permissive terms for commercial and research use.

Limitations

• No public information on the number of audio clips, speakers, or total hours of speech.
• Lack of documentation on recording conditions, speaker demographics, or transcript validation introduces uncertainty.

Provenance

Source

neongeckocom on Hugging Face.

Collection Method

Inferred to be a curated collection of clean speech recordings, likely for TTS model training.

Freshness

Last updated on the Hugging Face platform in September 2022.

Geography

Region tag suggests United States as a primary source, but coverage is not confirmed.