MOSS-TTS vs Chatterbox: which open TTS model to use?
Updated By Agent Radar6 min read
Short answer
MOSS-TTS vs Chatterbox: MOSS-TTS is an 8B-flagship family for long-form, dialogue and 31 languages; Chatterbox is a small MIT family for voice agents.
Top pick: MOSS-TTS. An Apache-2.0 speech and sound generation family from MOSI.AI and the OpenMOSS team, with separate models for long-form speech, multi-speaker dialogue, voice design, sound effects and realtime streaming.
Pick Chatterbox if you want a small, MIT-licensed model that clones a voice from a short clip and runs on a modest GPU or a CPU, especially for English voice agents. Pick MOSS-TTS if you need more languages, long-form narration, multi-speaker dialogue or fine pronunciation control, and have the GPU for an 8B flagship. Both families are permissively licensed and both now ship a CPU-sized model, so the split is scale and scope. Sources checked 2026-10-28.
Quick comparison
| MOSS-TTS | Chatterbox | |
|---|---|---|
| Maker | MOSI.AI and the OpenMOSS team | Resemble AI |
| License | Apache-2.0 | MIT |
| Flagship size | 8B (MOSS-TTS-v1.5) | 500M (Multilingual V3) |
| Smallest model | MOSS-TTS-Nano, about 100M | Chatterbox-Nano, 110M |
| Languages | 31 in v1.5 | 23 in Multilingual V3; Turbo and Nano English only |
| Voice cloning | Zero-shot | Zero-shot, from a reference clip |
| Realtime model | MOSS-TTS-Realtime, 1.7B, 180 ms TTFB after warm-up | Chatterbox-Turbo, 350M, one-step decoder |
| Dialogue | MOSS-TTSD for multi-speaker dialogue | Not a separate model |
| Expressive control | Pause markers, Pinyin, IPA, duration | Paralinguistic tags, exaggeration and CFG weight |
| Watermark | None described | PerTh watermark on every output |
| Install | Clone the repo; Python 3.12, torch 2.9.1+cu128 | pip install chatterbox-tts (PyPI 0.1.7) |
How they differ
Scope
MOSS-TTS is a family of separate models, each for a job. The README lists MOSS-TTS for long-form speech and cloning, MOSS-TTSD for multi-speaker dialogue and podcasts, MOSS-VoiceGenerator for designing voices from text, MOSS-SoundEffect for environmental audio and MOSS-TTS-Realtime for voice agents. The latest v1.5 flagship adds language tags, more stable cloning and inline pause markers such as [pause 3.2s].
Chatterbox is narrower: text-to-speech with zero-shot cloning, in several sizes. Multilingual V3 is the general model; the README says it keeps the 0.5B size of V2 while improving speaker similarity and reducing hallucinated continuations. A Single Language Pack adds six finetunes for Chinese, Latin American and European Spanish, Brazilian and European Portuguese, and Hindi.
Size and hardware
This is the biggest practical difference. The MOSS-TTS-v1.5 flagship is 8B parameters, and the local-transformer v1.5 is 4B. The README notes that the 8B model fits on 8GB GPUs through its llama.cpp pipeline with GGUF weights; the standard install targets CUDA 12.8 wheels.
Chatterbox's largest model is 500M. Turbo is 350M, and the README says its distilled decoder needs one step instead of ten, cutting compute and VRAM. Nano, at 110M, targets on-device and CPU use.
Voice agents and latency
MOSS-TTS-Realtime is built for multi-turn agents and models prior context and user audio. The README reports a 180 ms time to first byte after warm-up and a real-time factor of 0.51.
Chatterbox's agent model is Turbo, English only, with paralinguistic tags such as [chuckle] for more natural replies. The README does not give a local latency figure for Turbo; the sub-200 ms figure it mentions is for Resemble's hosted service.
Languages
MOSS-TTS-v1.5 lists 31 languages: the 20 from 1.0 plus Cantonese, Dutch, Finnish, Hindi, Macedonian, Malay, Romanian, Swahili, Tagalog, Thai and Vietnamese. Chatterbox Multilingual V3 lists 23, including Norwegian, which MOSS-TTS does not list. Check both lists for your languages before choosing.
Watermarking and responsible use
Chatterbox embeds Resemble's PerTh watermark in every output and documents how to detect it. That helps provenance; it also means generated audio is marked. The MOSS-TTS README does not describe a watermark.
Which one to choose
- English voice agent on a single consumer GPU or CPU: Chatterbox Turbo or Nano.
- Multilingual agent with streaming: MOSS-TTS-Realtime, or Chatterbox Multilingual V3 if your languages are on its list.
- Audiobooks and long narration: MOSS-TTS-v1.5, which is built for stable long-form speech.
- Podcasts or multi-speaker dialogue: MOSS-TTSD.
- Quick prototype with the fewest dependencies: Chatterbox, installed from PyPI.
Who should skip each
Skip MOSS-TTS if you want a one-line install and a small model: apart from Nano, its models run from 1.7B to 8B parameters, and the standard install pins CUDA 12.8 torch wheels. Skip Chatterbox if you need languages beyond its 23, multi-speaker dialogue in one pass, or output without a watermark.
How we compared
We read the OpenMOSS/MOSS-TTS README, the resemble-ai/chatterbox README, the ResembleAI/chatterbox Hugging Face model card, GitHub repository metadata for both projects and the chatterbox-tts PyPI metadata, all on 2026-10-28. We did not run either model, listen to samples or benchmark speed, quality or memory use. Latency and speed figures are the projects' own.
Limitations
The cited sources contain no head-to-head benchmark of MOSS-TTS and Chatterbox. MOSS-TTS evaluations in its README compare against other systems, and Chatterbox's Podonos evaluations compare Turbo with hosted services. Hosted pricing for either is not stated here.
Related: MOSS-TTS hub, Chatterbox hub, MOSS-TTS vs Kokoro, Kokoro alternatives, the voice models category and Best free AI voice models in 2026.
The tools
An Apache-2.0 speech and sound generation family from MOSI.AI and the OpenMOSS team, with separate models for long-form speech, multi-speaker dialogue, voice design, sound effects and realtime streaming.
Pros
- MOSS-TTS-v1.5 lists 31 languages, with language tags and explicit pause markers
- Dedicated models for dialogue (MOSS-TTSD), voice design and realtime agents
- Pinyin, IPA and token-level duration control
- Serving options documented for vLLM-Omni, SGLang-Omni, llama.cpp with GGUF weights, and mlx-audio
Cons
- Flagship MOSS-TTS-v1.5 is an 8B model; the README's llama.cpp note says it fits on 8GB GPUs only through that pipeline
- Setup pins torch 2.9.1+cu128 and recommends Transformers 5.0.0
- No built-in output watermark is described in the README
Resemble AI's MIT-licensed TTS family: a 500M Multilingual V3 model for 23 languages, a 350M English Turbo model for voice agents and a 110M Nano model for CPUs, all with zero-shot voice cloning.
Pros
- Small models: 500M multilingual, 350M Turbo, 110M Nano
- Nano runs 3x faster than realtime on 8 CPU cores, per the README
- Paralinguistic tags such as [laugh] and [cough] in Turbo and Nano
- One-line install: pip install chatterbox-tts
Cons
- 23 languages in the general multilingual model, fewer than MOSS-TTS-v1.5
- Turbo and Nano are English only
- Every output carries Resemble's PerTh watermark, per the README
Questions
Is MOSS-TTS or Chatterbox better for voice agents?
For English agents on modest hardware, Chatterbox Turbo (350M) or Nano (110M) is the lighter choice. For multilingual agents, MOSS-TTS-Realtime is a 1.7B model built for multi-turn streaming, with a reported 180 ms time to first byte after warm-up.
Which supports more languages, MOSS-TTS or Chatterbox?
MOSS-TTS. MOSS-TTS-v1.5 lists 31 languages, including Cantonese, Thai and Vietnamese. Chatterbox Multilingual V3 lists 23, plus six single-language finetunes such as Latin American Spanish and Brazilian Portuguese.
Can MOSS-TTS and Chatterbox run on a CPU?
Both have a small CPU model. MOSS-TTS-Nano is about 100M parameters and streams on 4 CPU cores, per the MOSS-TTS README. Chatterbox-Nano is 110M parameters and runs 3x faster than realtime on 8 CPU cores, per the Chatterbox README.
Are MOSS-TTS and Chatterbox free for commercial use?
Both are under permissive licenses: the MOSS-TTS README says its models are Apache-2.0, and the Chatterbox repository and model card list MIT. Check the card of the exact checkpoint you download.
Does Chatterbox watermark its audio?
Yes. The README says every file generated by Chatterbox includes Resemble AI's PerTh neural watermark, and gives a script to detect it. The MOSS-TTS README does not describe a watermark.