MOSS-TTS

MOSS-TTS covers OpenMOSS's 8B MossTTSDelay v1.5 and v1.0 models plus the local-transformer variants. The MLX port uses the upstream Qwen3 backbone weights, RVQ generation path for each variant, and the matching MOSS Audio Tokenizer.

Quick Start

from mlx_audio.audio_io import write as audio_write
from mlx_audio.tts import load

model = load("OpenMOSS-Team/MOSS-TTS-v1.5", lazy=True)

result = next(model.generate(
    text="Hello, this is MOSS-TTS running on MLX.",
    max_tokens=120,
))
audio_write("moss_tts.wav", result.audio, result.sample_rate)

For the local-transformer variant, load:

model = load("OpenMOSS-Team/MOSS-TTS-Local-Transformer", lazy=True)

For the v1.5 local-transformer checkpoint, load:

model = load("OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5", lazy=True)

The v1.5 local checkpoint uses OpenMOSS-Team/MOSS-Audio-Tokenizer-v2, 48 kHz output, and fixed 12-codebook RVQ generation.

The v1.5 local checkpoint also supports streamed generation:

chunks = model.generate(
    text="Hello, this is streamed MOSS local transformer audio.",
    language="English",
    stream=True,
    streaming_interval=2.0,
)

The first chunk is emitted after 4 generated RVQ frames by default. Subsequent chunks use streaming_interval seconds of generated frames. MLX keeps a single-stream MOSS-Audio-Tokenizer-v2 decoder session during streaming decode.

For non-English v1.5 prompts, pass language when known:

result = next(model.generate(
    text="Bonjour, je voudrais essayer une voix francaise naturelle.",
    language="French",
    max_tokens=120,
))

The v1.5 checkpoints accept inline pause markers such as [pause 3.2s].

Voice Cloning

result = next(model.generate(
    text="This is a short cloned voice sample.",
    ref_audio="speaker.wav",
    max_tokens=120,
))
audio_write("moss_tts_clone.wav", result.audio, result.sample_rate)

Notes

• The delay-pattern models use 24 kHz audio. The v1.5 local-transformer model uses 48 kHz stereo audio.
• The full audio tokenizer is required for reference encoding and waveform decoding.