Overview
MOSS-TTSD is an open-source speech model for long conversations with multiple speakers. Give it a script with speaker labels, reference recordings, and their transcripts. The model uses dialogue context and previously generated audio to shape speaker transitions, timing, and delivery, keeping each voice recognizable as the exchange develops.
Built on MOSS-TTS, it is trained for spoken formats including interviews, sports and esports commentary, audiobooks, dubbing, and Chinese crosstalk. It can capture exchanges between a host and guest, passages that alternate narration and dialogue, and commentary that shifts from analysis to excited calls. These different speaking styles can unfold within one continuous performance.
Capabilities
A cast of up to five speakers
Generate one conversation with one to five speakers. Speaker labels assign each line, and individual reference recordings establish the voices, so hosts, guests, narrators, and characters can share the same script.
Natural turns and overlapping voices
The model accounts for the timing of replies, interjections, and overlapping voices. Context shapes how speakers move between turns, giving the exchange a conversational rhythm while preserving distinct voices.
Up to 60 minutes in one generation
Generate up to 60 minutes of continuous speech in a single pass. Long-context modeling carries voices and prosody forward as topics and speakers change, supporting extended interviews, full episodes, and audio stories.
Set each voice with a short recording
A 3–10-second reference recording and its transcript are enough for zero-shot voice cloning, with no speaker-specific training. Assign a reference to each speaker, then use those voices throughout the dialogue.
Voices across 20 languages
Create dialogue in 20 languages, including Chinese, English, Japanese, and Korean. Cross-language voice cloning lets you use a reference voice for lines in another supported language, keeping the character’s vocal identity.
From measured analysis to excited calls
The model supports the shifts in delivery found in sports and esports commentary, from measured analysis to excited calls. Training across podcasts, dubbing, and other spoken formats also covers the different demands of narration and character dialogue.