ParsVoice is a Persian speech and text corpus for training multi-speaker text-to-speech systems. It holds about 1.36 million audio segments totaling roughly 2,178 hours.
DatasetFree
Step 1: Sign in to Hugging Face and accept the access conditions on the dataset page.
Step 2: Load the default config with the datasets library, using streaming mode because the full corpus is about 130 GB.
Step 3: Read the transcript, speaker_id and audio fields from each sample.
Step 4: Load the metadata config on its own to inspect quality statistics without downloading audio.
Step 5: Filter by completion status and audio quality score before training, then join back to the audio.
CC-BY-NC-4.0, which is non-commercial. The files also sit behind a gate where you accept the conditions.
It has 1,360,521 segments totaling about 2,178 hours. The full audio is about 130 GB, so the card recommends streaming.
Speaker IDs are inferred automatically by clustering. Segments marked as having a known narrator are corroborated by source metadata, if you need labels that are not purely automatic.
Yes. The metadata config holds the per-segment quality fields and is about 70 MB.