niobures/SASPEECH_AUTO_clean

🎙️ SASPEECH (Automatic) - Cleaned & IPA Transcribed

Dataset Description

This dataset is a highly refined, TTS-ready version of the SASPEECH "automatic" subset. It features the voice of prominent Israeli journalist and podcaster Shaul Amsterdamsky (from the Chayot Kis podcast).

Unlike the original release, this version has been strictly cleaned and pre-processed specifically for modern Text-to-Speech (TTS) training. Crucially, the text transcripts have been converted into the International Phonetic Alphabet (IPA) to bypass the ambiguities of unvowelized Hebrew, making it plug-and-play for acoustic models.

• Language: Hebrew (he)
• Primary Speaker: Shaul Amsterdamsky (Male)
• Transcription Format: IPA (International Phonetic Alphabet)
• Audio Quality: Cleaned and normalized for TTS (e.g., 22.05kHz / 24kHz / 44.1kHz)

Uses

This dataset is designed to solve the "Hebrew TTS bottleneck"—the lack of diacritization (Niqqud) in standard text.

Primary use cases include:

1. End-to-End TTS Training: Train models directly on IPA phonemes for highly accurate, stumble-free Hebrew pronunciation.
2. Acoustic Modeling: Perfect for training fast, modern architectures (like DiffMamba, VITS, or FastSpeech2) that require explicit phonetic alignment.
3. Voice Cloning: A robust, single-speaker phonetic baseline for few-shot voice conversion.