MOSS-TTS-Realtime - Streaming Text-to-Speech with Voice Cloning

Overview

MOSS-TTS-Realtime is the streaming model in the MOSS-TTS family from OpenMOSS and MOSI.AI. OpenMOSS built it for voice agents: audio comes back in small chunks while the rest of the passage is still being synthesized, so playback can start before the reply is finished. OpenMOSS lists it at 1.7B parameters (the published checkpoint holds about 2.3B), and it uses OpenMOSS's MOSS-Audio-Tokenizer codec at 24 kHz to turn speech into tokens and back.

The model conditions each reply on the text and audio of earlier dialogue turns, which keeps one voice consistent across a conversation. OpenMOSS lists a maximum context length of 32K, about 40 minutes by its estimate. Training used more than 2.5 million hours of single-speaker speech and over 1 million hours of two-speaker and multi-speaker conversation.

On Seed-TTS-eval, OpenMOSS reports a 1.971% word error rate and 68.9 speaker similarity for English, and a 1.07% character error rate and 76.7 similarity for Chinese.

What you can build

• Spoken replies for a voice assistant or support bot that start playing before the whole answer has been synthesized
• Narration in a cloned voice for videos, audiobook chapters or podcast segments
• Read-aloud buttons for articles and app screens in any of the 20 supported languages
• Character voices for games and interactive stories, each cloned from its own reference clip

Inputs

• Text: the passage to speak. There is no language selector, so write it in the language you want to hear: Chinese, English, German, Spanish, French, Japanese, Italian, Hebrew, Korean, Russian, Persian, Arabic, Polish, Portuguese, Czech, Danish, Swedish, Hungarian, Greek or Turkish.
• Reference Audio (optional): one clip of the voice to clone, as MP3, WAV, M4A or OPUS. Without a clip, nothing is cloned and the model chooses the voice itself.

Advanced settings control sampling. Temperature (0 to 2, default 0.8) trades stability for expressiveness. Top P (default 0.6) and Top K (default 30) narrow which audio tokens can come next. Repetition Penalty (default 1.1) discourages repeated audio patterns within the last Repetition Window tokens (default 50).

Outputs

Speech streams back as 24 kHz mono audio chunks while generation continues. You can end a session early and keep what was spoken up to that point. When the session closes, all of the audio from that run is saved as one MP3 file.

Recommended settings

The defaults are the decoding values OpenMOSS recommends for this model. For cloning, use a clean recording of one speaker without music or background noise. If the voice drifts or words come out garbled, lower Temperature or Top P a little. If a sound or syllable keeps repeating, raise Repetition Penalty slightly.

Limitations

• OpenMOSS publishes evaluation numbers for English and Chinese only. The other 18 languages have no reported scores.
• Each run is one turn with a fresh context, so the model's multi-turn memory does not carry from one run to the next. Reuse the same reference clip to keep a voice consistent across runs.

Safety & compliance

MOSS-TTS-Realtime is released under the Apache 2.0 license. Clone only voices you have permission to use, and label synthetic speech wherever listeners could mistake it for a real recording.