MOSS-TTS v1.5 Multilingual Voice Cloning

Overview

MOSS-TTS-v1.5 is an autoregressive text-to-speech model by OpenMOSS and MOSI.AI. It generates discrete audio tokens, then decodes them into speech audio. It supports zero-shot voice cloning from a short reference clip, so you can keep a consistent speaker without training. It also adds explicit pause control and stronger multilingual stability when you provide a language tag.

What you can build

• Voiceovers for product demos, explainers, and slides
• Long-form narration, including multi-minute and extended reads
• Multilingual IVR prompts and support content across 31 languages
• Voice cloning for consistent character voices in apps and games
• Pronunciation-controlled learning content using Pinyin or IPA hints
• Scripted speech with explicit, timed pauses for pacing

Inputs

• Text to synthesize, provided as plain text. It can mix normal text with Pinyin or IPA for pronunciation control.
• Optional inline pause markers embedded in the text, using the form [pause X.Ys].
• Optional language tag you select to guide multilingual synthesis. This is recommended for languages beyond Chinese and English.
• Optional reference audio clip for zero-shot voice cloning. Provide a single clip when you want speaker similarity.
• Optional target duration control, provided as an expected count of audio tokens. The model uses about 12.5 audio tokens per second.
• Optional generation length cap, set as a maximum number of generated tokens. Raise it for long passages.
• Optional decoding controls for text planning, such as temperature and sampling filters. Lower values make results more deterministic.
• Optional decoding controls for acoustic generation, such as temperature and sampling filters. Higher values can add variety but may reduce stability.
• Optional anti-repetition control for audio tokens. Slight increases can reduce stutters and loops.

Outputs

The model returns synthesized speech audio as a WAV result. The audio content is the spoken rendition of your input text, including any explicit pauses you inserted. If you provide reference audio, the output aims to match the reference speaker’s timbre and some speaking style. When you use duration control, the output length shifts toward your target token count.

Recommended settings

• General TTS: use an audio temperature of 1.7, top-p of 0.8, top-k of 25, and an audio repetition penalty of 1.0.
• Multilingual text: always set the language tag when you know it.
• Long passages: increase the generation token cap so the model can finish the full script.
• Voice cloning: use a clean reference clip with one speaker and minimal background noise.

Limitations

• The model is sensitive to decoding settings. Aggressive sampling can cause instability or artifacts.
• Voice cloning quality depends heavily on the reference clip. Noise, music, or multiple speakers reduce similarity.
• Very long texts need a higher generation token cap. Otherwise, audio may stop early.
• Language tags matter for non-English and non-Chinese text. Missing tags can reduce stability.
• Messy inputs can hurt prosody. This includes unpunctuated text, inconsistent casing, or malformed pause markers.

Safety & compliance

Models in the MOSS-TTS family are released under the Apache License 2.0.

Voice cloning can enable impersonation. Only clone voices you own or have explicit permission to use. Don’t use generated audio to mislead people about identity, consent, or endorsements.