SoulX-Podcast is an inference codebase for generating high-fidelity podcasts from text with support for multiple speakers, dialects, and paralinguistic features.
The tool addresses the challenge of creating realistic long-form podcast audio by combining multi-speaker dialogue generation with fine-grained control over speech characteristics. It handles multi-turn conversational scenarios where different speakers alternate, moving beyond traditional single-speaker text-to-speech synthesis. The system supports Mandarin and English as primary languages, along with several Chinese dialects including Sichuanese, Henanese, and Cantonese. Users can inject paralinguistic events such as laughter, sighs, breathing, coughing, and throat clearing through special tags to enhance naturalness. The tool also enables zero-shot voice cloning, allowing speakers to be defined from reference audio samples without requiring dialect-specific training data.
This tool suits projects requiring podcast-style audio generation, particularly those targeting multilingual or multi-dialect audiences. It is most valuable for creators building conversational audio content where speaker diversity and natural speech patterns matter. The zero-shot cross-dialectal voice cloning capability distinguishes it from standard text-to-speech systems, as it allows generating speech in dialects different from the reference speaker's native dialect. Teams working with Mandarin or English content, especially those needing dialectal variation, will find the dialect support particularly relevant.
The project maintains an active demo page and provides model weights through a public collection. A technical report documents the approach and performance characteristics. The codebase is organized as inference code rather than a training framework, indicating focus on deployment and practical use of pre-trained models.