MOSS TTSD
A local text-to-speech model generates long conversations with distinct speakers, voice cloning and multilingual dialogue under Apache 2.0.
MOSS TTSD generates spoken dialogue from scripts for podcasts, audio dramas, dubbing and commentary. Its focus is maintaining distinct voices across long conversations, including speaker changes and overlapping speech.
- Speaker control: Assign dialogue to up to five speakers while keeping each voice consistent across the recording.
- Long conversations: Generate up to an hour of audio in a single session, with support for conversational turn-taking.
- Voice cloning: Reproduce a speaker's voice from a short reference recording and its matching transcript, without separate training for that person.
- Multilingual dialogue: Generate speech and clone voices across languages including Chinese, English, Japanese, German, Spanish and Arabic.
- Dialogue evaluation: TTSD-eval measures speaker attribution and voice similarity for developers comparing generated conversations.
MOSS TTSD is open source under the Apache 2.0 license. Local inference uses PyTorch and Hugging Face Transformers with MOSS-Audio-Tokenizer, and the supplied example supports CPU execution or NVIDIA GPUs through CUDA. SGLang is an alternative backend for accelerated inference.
