OpenMOSS / MOSS-TTS-Realtime
MOSS-TTS-Realtime
Real-time streaming text-to-speech with zero-shot voice cloning. Supports 20 languages including English, Chinese, Japanese, Korean, and more. Audio starts playing immediately — no waiting for full generation. Clone any voice from a short reference clip.
## Overview
MOSS-TTS-Realtime is a context-aware, real-time streaming text-to-speech model from the OpenMOSS team. Unlike conventional TTS that generates the entire audio before playback, MOSS-TTS-Realtime starts streaming audio as it generates — you hear the speech within milliseconds, even for long paragraphs. It supports zero-shot voice cloning from a short reference audio, producing natural and expressive speech across 20 languages.
## What you can build
- Voice agents and real-time conversational AI
- Audiobook and podcast narration with custom voices
- Live dubbing and voice-over tools
- Multilingual content creation
- Accessibility tools for text-to-speech
- Voice cloning applications with short audio references
## Key features
- Real-time streaming: Audio starts playing immediately as it generates — no waiting for the full text to be synthesized
- Zero-shot voice cloning: Clone any voice from a short reference audio clip (wav, mp3, flac, ogg, m4a, wma, opus)
- 20 languages: Chinese, English, German, Spanish, French, Japanese, Italian, Hebrew, Korean, Russian, Persian, Arabic, Polish, Portuguese, Czech, Danish, Swedish, Hungarian, Greek, Turkish
- Context-aware synthesis: Models dialogue history across text and acoustics for natural multi-turn conversations
- Long-form stability: Supports up to 32K context (~40 minutes) with consistent voice and prosody
- High expressiveness: Trained on 2.5M+ hours of speech data for highly natural prosody
- 2.3B parameters: Powerful yet efficient architecture based on Qwen3 language model backbone
## Inputs
- Text: The text to synthesize into speech (required)
- Reference Audio: A short audio clip of the voice to clone (optional). If not provided, the model uses its default voice. Supported formats: wav, mp3, flac, ogg, m4a, wma, opus
## Outputs
- Streaming audio: Real-time PCM audio stream at 24kHz — playback begins immediately
- Saved audio file: Complete audio saved as MP3 (or WAV fallback) when the session ends
## How it works
1. Click Run and enter your text
2. Optionally upload a reference audio for voice cloning
3. Audio starts streaming immediately — you hear it as it generates
4. Click End Session to stop early, or let it finish naturally
5. Download the complete audio file after the session ends
## Recommended settings
- For voice cloning, use a clear 5-15 second reference clip with minimal background noise
- Long texts work great — the model streams audio progressively without delay
- The model auto-detects language from the input text
## Limitations
- Voice cloning quality depends on the clarity and length of the reference audio
- Very short reference clips (under 2 seconds) may produce lower quality clones
- Some languages may have lower quality than Chinese and English
- Session ends automatically after the full text is synthesized
## Technical details
- Architecture: Qwen3-based language model + local transformer for audio token generation
- Audio codec: MOSS-Audio-Tokenizer at 24kHz
- Parameters: 2.3B (BF16 precision)
- Training data: 2.5M+ hours single-speaker + 1M+ hours multi-speaker conversational
- License: Apache 2.0
- Source: [OpenMOSS-Team/MOSS-TTS-Realtime](https://huggingface.co/OpenMOSS-Team/MOSS-TTS-Realtime)
API quick start
Run MOSS-TTS-Realtime with a single API call.
{
"text": "A small robot learned to paint by watchin…",
"inputAudio": "https://your-cdn.com/input.mp3",
"temperature": 0.8,
"topP": 0.6
}