MOSS-TTS Documentation

MOSS-Audio-Tokenizer is the unified discrete audio interface for the MOSS-TTS family (including MOSS-TTS, MOSS-TTSD, MOSS-VoiceGenerator, MOSS-SoundEffect, and MOSS-TTS-Realtime). It is based on the Cat (Causal Audio Tokenizer with Transformer) architecture.

Key features:

• Unified Backbone: Provides a consistent audio representation across the MOSS ecosystem.
• High Compression: Compresses 24kHz raw audio to a 12.5Hz frame rate using a 32-layer Residual Vector Quantizer (RVQ).
• Variable Bitrate: Supports reconstruction from 0.125kbps to 4kbps.
• Streaming Native: Designed for low-latency streaming inference via a pure Causal Transformer architecture.

For detailed setup and evaluation metrics, refer to the MOSS-Audio-Tokenizer Repository.

MOSS-TTS-Realtime is a low-latency speech synthesis architecture designed for real-time conversational scenarios. It uses a hierarchical input design where text and speech tokens are processed at different levels, supporting streaming text input and streaming audio output.

Key components include:

• A 1.7B-parameter backbone (initialized from Qwen3-1.7B) that encodes linguistic and contextual information.
• A 200M-parameter local Transformer that uses the backbone's hidden states to autoregressively generate RVQ-based audio tokens.
• MOSS-Audio-Tokenizer for audio reconstruction and high-fidelity waveform synthesis.

MOSS-TTS-Nano is a lightweight TTS model optimized for CPU-first and real-time deployment scenarios. It uses an Audio Tokenizer + LLM pipeline (purely autoregressive) to provide low-latency streaming generation without requiring a GPU.

Key Features:

• 0.1B Parameters: Compact model size for low memory footprint and deployment costs.
• CPU Optimized: Capable of real-time streaming generation on as little as a 4-core CPU.
• Multi-lingual Voice Clone: Supports cross-lingual synthesis using a single reference audio clip.
• High Fidelity: Supports 48 kHz stereo input and output.

For detailed environment configuration, advanced usage, and evaluation metrics, visit the MOSS-TTS-Nano repository.

MOSS-TTS is a production-grade TTS foundation model designed for high-quality speech synthesis. Key features include:

• Zero-shot Voice Cloning: Clone a speaker's timbre and style from short reference audio without fine-tuning.
• Ultra-long Speech Generation: Supports continuous generation for up to one hour.
• Fine-grained Control:
  • Token-level Duration Control: Adjust pacing, rhythm, and pauses at the token level.
  • Phoneme-level Pronunciation Control: Supports pure Pinyin, pure IPA, or mixed Chinese/English/Pinyin/IPA input.
• Multilingual & Code-switching: Supports 31 languages and natural mixed-language generation (e.g., Chinese–English) within a single utterance.
• Explicit Pause Markers: Supports inline markers like [pause 3.2s] (v1.5+).

MOSS-TTS Family is an open-source speech and sound generation model family developed by MOSI.AI and the OpenMOSS team. It is designed for high-fidelity, high-expressiveness, and complex real-world scenarios. Key capabilities include:

• Stable long-form speech
• Multi-speaker dialogue
• Voice and character design
• Environmental sound effects
• Real-time streaming TTS

MOSS-Audio-Tokenizer is the unified discrete audio interface for the MOSS-TTS family (including MOSS-TTS, MOSS-TTSD, MOSS-VoiceGenerator, MOSS-SoundEffect, and MOSS-TTS-Realtime). It is based on the Cat (Causal Audio Tokenizer with Transformer) architecture.

Key Features:

• Unified Representation: Provides a shared backbone for consistent audio representation across the MOSS family.
• High Compression & Fidelity: Compresses 24kHz raw audio to a 12.5Hz frame rate using 32-layer Residual Vector Quantization (RVQ), supporting variable bitrates from 0.125kbps to 4kbps.
• Native Streaming Design: Built with a pure Causal Transformer architecture for low-latency streaming inference.

For configuration and advanced usage, visit the MOSS-Audio-Tokenizer repository.

Several community-driven integrations and tools are available for the MOSS-TTS ecosystem:

• ComfyUI-MOSS-TTS: An extension for ComfyUI.
• MOSS-TTS-OpenAI: An OpenAI-compatible TTS API for MOSS-TTS.
• AnyPod: A podcast generation tool using MOSS-TTS/MOSS-TTSD as the backend.
• Norwegian LoRA: A community-trained LoRA adapter (mlp, r=16) for Norwegian speech, available on Hugging Face.

MOSS-TTS provides two architectural patterns to balance performance and latency:

Architecture A: Delay Pattern (MossTTSDelay)

• Mechanism: Single Transformer backbone with (n_vq + 1) heads using delay scheduling for multi-codebook audio tokens.
• Best For: Production environments requiring strong long-context stability, efficient inference, and stable long-form synthesis.

Architecture B: Global Latent + Local Transformer (MossTTSLocal)

• Mechanism: Produces a global latent per time step, with a lightweight Local Transformer emitting token blocks.
• Best For: Streaming-friendly applications and research/evaluation where objective benchmarks are a priority.