MOSS-TTS

Overview

MOSS-TTS is an open-source speech synthesis foundation model from the OpenMOSS team, released on February 6, 2026 — the date of the initial commit and the model weights init commit on the OpenMOSS-Team/MOSS-TTS Hugging Face repository. The public GitHub project and README followed on February 10, 2026. A refresh, MOSS-TTS-v1.5, was published on May 25, 2026.

OpenMOSS describes the wider release as "an open-source speech and sound generation model family from MOSI.AI and the OpenMOSS team," designed for "high-fidelity, high-expressiveness, and complex real-world scenarios." Within that family, MOSS-TTS is positioned as "the flagship base model" and "a next-generation, production-grade TTS foundation model focused on voice cloning, ultra-long stable speech generation, token-level duration control, multilingual & code-switched synthesis, and fine-grained Pinyin/phoneme-level pronunciation control."

The architecture is deliberately conventional. The MOSS-TTS Technical Report (arXiv 2603.18090, submitted March 18, 2026) frames it as "a speech generation foundation model built on a scalable recipe: discrete audio tokens, autoregressive modeling, and large-scale pretraining." Everything rests on MOSS-Audio-Tokenizer, described in the abstract as "a causal Transformer tokenizer that compresses 24 kHz audio to 12.5 fps with variable-bitrate RVQ and unified semantic-acoustic representations." At 12.5 frames per second, one second of audio costs roughly 12.5 tokens — which is what makes hour-long generation tractable for an autoregressive transformer.

Both the weights and the code are Apache 2.0, with no revenue cap, field-of-use restriction, or geographic carve-out. Training data is not released.

A note on naming

Three claims about MOSS-TTS circulate widely and are wrong:

• It is not the most-downloaded Chinese text-to-speech model on Hugging Face. Queried on July 8, 2026, OpenMOSS-Team/MOSS-TTS reports 1,079,657 downloads over the previous 30 days — sixth among all text-to-speech-tagged models, behind Kokoro-82M (13,509,796), XTTS-v2 (9,828,098), Chatterbox (2,463,153), and two Alibaba checkpoints: Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice (2,098,492) and Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice (1,319,373). MOSS-TTS is therefore the third most-downloaded Chinese TTS checkpoint in that tagged ranking. Alibaba's five Qwen3-TTS model repositories together report 6,755,528 downloads — roughly 6.3× MOSS-TTS's figure. Note that one of them, Qwen/Qwen3-TTS-12Hz-1.7B-Base (2,128,870), carries no pipeline_tag at all and so never appears in the tagged ranking; on raw downloads it would sit fourth.
• There is no MOSS-TTS v2. The newest checkpoint is v1.5. OpenMOSS-Team/MOSS-TTS-v2 and OpenMOSS-Team/MOSS-TTS-v1.6 do not resolve on the Hugging Face API.
• MOSS-TTS did not "beat Gemini." The family README attributes that result to MOSS-TTSD, the dialogue model, which it says "outperformed top closed-source models like Doubao and Gemini 2.5-pro" in subjective evaluations. MOSS-TTS carries no such claim.

Capabilities

• Zero-shot voice cloning: Clone "a target speaker's timbre (and part of speaking style) from short reference audio, without speaker-specific fine-tuning." OpenMOSS says v1.5 "improves speaker similarity and reduces cloning variance."
• Hour-long stable generation: The model card claims "continuous long-form speech generation for up to one hour in a single run," holding speaker identity across the span.
• Token-level duration control: A tokens argument sets the expected audio length at a stated rate of "1s ≈ 12.5 tokens," giving direct control over pacing and pauses.
• Pinyin and IPA pronunciation control: Input accepts raw text, tone-numbered Pinyin (nin2 hao3), IPA wrapped in slashes (/həloʊ/), or any mixture of Chinese, English, Pinyin, and IPA in a single string.
• Multilingual and code-switched synthesis: 20 languages in 1.0, 31 in v1.5, with mixed-language utterances rendered "with smooth transitions, consistent speaker identity, and pronunciation-aware rendering on both sides of the switch."
• Explicit pause markers (v1.5 only): Inline [pause 3.2s] tokens insert a timed silence, alongside "more stable punctuation-following prosody."
• Two published architectures: An 8B delay-pattern model for production and a 1.7B local-transformer model for research, trained under one recipe as "reproducible references for deployment and research."

Technical Specifications

• Hugging Face repositories: OpenMOSS-Team/MOSS-TTS (1.0) and OpenMOSS-Team/MOSS-TTS-v1.5
• Parameters: 8B per the model card; the Hugging Face safetensors index reports 8,489,841,664 parameters in BF16 for both 1.0 and v1.5
• Architecture: MossTTSDelay — a single autoregressive transformer backbone with (n_vq + 1) heads, using delay scheduling over multi-codebook audio tokens
• Companion architecture: MossTTSLocal (labelled 1.7B) — global latent per timestep plus a lightweight local transformer emitting a token block per step; streaming-friendly, no delay scheduling
• Audio tokenizer: MOSS-Audio-Tokenizer, built on the "Cat" (Causal Audio Tokenizer with Transformer) architecture — a "CNN-free" stack of causal transformer blocks, "scaling to 1.6 billion parameters"
• Tokenizer compression: 24 kHz audio → 12.5 Hz frame rate, via a 32-layer Residual Vector Quantizer supporting 0.125–4 kbps
• Trained from scratch: The tokenizer "does not rely on any pretrained encoders (such as HuBERT or Whisper) or distillation from teacher models"
• Languages: 20 (1.0); 31 (v1.5)
• Runtime: Transformers 5.0.0, trust_remote_code=True, optional FlashAttention 2
• Recommended decoding: audio_temperature 1.7, audio_top_p 0.8, audio_top_k 25 for MossTTSDelay-8B
• License: Apache 2.0 (weights and code)

Size labels do not always match the weights. MOSS-TTS's own "8B" label agrees with its safetensors index (8,489,841,664). The sibling checkpoints do not: MOSS-TTS-Local-Transformer is labelled 1.7B but indexes 3,060,606,464 parameters, MOSS-TTS-Realtime is labelled 1.7B but indexes 2,331,940,864, MOSS-VoiceGenerator is labelled 1.7B but indexes 2,114,118,656, and MOSS-Audio-Tokenizer is described as "1.6B combined parameters (Encoder + Decoder)" but indexes 1,774,566,400. OpenMOSS does not explain the gap — the labels appear to be nominal backbone sizes, not total parameter counts. This page reports both figures rather than picking one.

OpenMOSS does not publish a training-data hour count, a knowledge cutoff, or VRAM requirements for MOSS-TTS, so this page does not state any.

Use Cases

• Long-form narration: Audiobooks, documentation read-alouds, and course material, where the one-hour single-run claim and identity stability matter more than latency.
• Voice cloning products: Zero-shot timbre transfer from a short reference clip, with no per-speaker training step.
• Multilingual and code-switched content: Chinese–English mixed utterances rendered by one model, with pronunciation controlled per-segment via Pinyin or IPA.
• Dubbing and timed captioning: Token-level duration control lets output be fitted to a target length, which cascade-free audio-processing pipelines otherwise handle with post-hoc time-stretching.
• Low-latency voice agents: Not MOSS-TTS itself — OpenMOSS points this workload at MOSS-TTS-Realtime, whose "TTFB (Time To First Byte)" it says "reaches 180 ms." That figure is vendor-reported and unaccompanied by a hardware description.
• Research baselines: Two open architectures over a shared tokenizer make the family a controlled setting for tokenization and discrete-audio-modeling ablations.

Performance / Benchmarks

OpenMOSS reports results on Seed-TTS-eval, an open zero-shot text-to-speech benchmark. Its headline vendor claim: MOSS-TTS "achieved state-of-the-art results on the open-source zero-shot TTS benchmark Seed-TTS-eval, not only surpassing all open-source models but also rivaling the most powerful closed-source models."

Selected rows from the model card's own table. Lower is better for WER/CER; higher is better for SIM (speaker similarity).

The vendor claim does not survive its own table. On intelligibility, two open-source models beat MOSS-TTS: Qwen3-TTS 1.7B posts the best English WER in the entire table (1.50 versus MossTTSDelay's 1.84), and GLM-TTS-RL posts the best open-source Chinese CER (0.89 versus 1.37). Where MOSS-TTS genuinely leads is speaker similarity — MossTTSLocal's 79.62 ZH SIM is the highest figure anywhere in the table, edging closed-source Seed-TTS at 79.6, and its 73.28 EN SIM leads all open models. Note also that the similarity lead belongs to the 1.7B research checkpoint, not the 8B production flagship.

Arena placement

MOSS-TTS does not appear on the Artificial Analysis Speech Arena leaderboard, nor on Artificial Analysis's text-to-speech model index. Fetched July 8, 2026, neither page contains a single occurrence of "MOSS," while Fish Audio and Qwen checkpoints are both listed. This is an absence, not a ranking: a model that is not submitted cannot be beaten.

For context, Artificial Analysis's own leaderboard copy on July 8, 2026 reads: "The top Text to Speech models by Elo rating are: 1. Simba 3.2 (Elo 1233), 2. Gemini 3.1 Flash TTS (Elo 1214), 3. Sonic 3.5 (Elo 1208), 4. Fun-Realtime-TTS (Elo 1204), 5. Realtime TTS 1.5 Max (Elo 1201)." Alibaba's closed, API-only Fun-Realtime-TTS topped that board when Artificial Analysis announced it on June 3, 2026 at Elo 1219; a month later it sits fourth.

Limitations

• Vendor-reported benchmarks, contradicted internally: Every Seed-TTS-eval number above is OpenMOSS's own. The "surpassing all open-source models" claim is not supported by the table printed directly beneath it. No independent replication was found.
• Not intelligibility-leading: On English WER, both Qwen3-TTS checkpoints in OpenMOSS's full table (1.50 and 1.68) beat MossTTSDelay's 1.84; only the stronger of the two appears in the selected rows above. On Chinese CER, ten of the seventeen open-source baselines in OpenMOSS's own table beat its 1.37 — led by GLM-TTS-RL (0.89) and VoxCPM (0.93). MOSS-TTS trades word-accuracy for timbre fidelity.
• The 8B flagship trails its own 1.7B sibling on similarity: MossTTSLocal wins both SIM columns (EN 73.28 vs 70.86; ZH 79.62 vs 76.98), while MossTTSDelay wins both intelligibility columns (EN WER 1.84 vs 1.93; ZH CER 1.37 vs 1.44). The split is two axes each, yet the card recommends the 8B "for production" and the 1.7B "for evaluation and research" on the grounds of its "SOTA objective metrics" — a recommendation the table only half supports.
• Sensitive to decoding hyperparameters: The model card warns that "MOSS-TTS is a pretrained base model and is sensitive to decoding hyperparameters." A temperature of 1.7 is unusually high and is not a safe default to vary blindly.
• No arena validation: Absent from Artificial Analysis's Speech Arena and model index entirely, so there is no blind human-preference Elo for it — only OpenMOSS's in-house preference testing.
• trust_remote_code=True required: Both checkpoints ship custom modeling code rather than an upstream Transformers architecture, which is a supply-chain consideration for production deployment.
• No published hardware, data, or safety documentation: No VRAM figures, no training-hour count, no voice-cloning consent or watermarking mechanism is described. An Apache 2.0 zero-shot cloning model with 3-second-class reference audio is a straightforward impersonation tool.
• Version drift in v1.5: OpenMOSS states that when the language field is omitted, v1.5 "may improve some languages and regress slightly on others compared with 1.0." The upgrade is only unambiguous if you tag the language explicitly.

Pricing & Access

Self-hosting is free. Both OpenMOSS-Team/MOSS-TTS and OpenMOSS-Team/MOSS-TTS-v1.5 are Apache 2.0. Commercial use, modification, and redistribution are permitted with no revenue ceiling and no registration requirement — a materially more permissive footing than the community licenses attached to many open-weights media models.

git clone https://github.com/OpenMOSS/MOSS-TTS.git
cd MOSS-TTS
pip install --extra-index-url https://download.pytorch.org/whl/cu128 -e .

Weights load through AutoModel / AutoProcessor with trust_remote_code=True on Transformers 5.0.0.

Hosted access is nominally offered through MOSI Studio (studio.mosi.cn), which the model card links for both an interactive demo and API documentation. Fetched on July 8, 2026, that site returns a maintenance banner — "系统维护中" ("The system is being upgraded") — and serves no rate card. OpenMOSS publishes no USD or CNY price for MOSS-TTS that could be verified, so this page states none.

Weights are also mirrored on ModelScope under the openmoss organization, and a GGUF conversion is published at OpenMOSS-Team/MOSS-TTS-GGUF.

Ecosystem & Tools

The MOSS-TTS Family is five task-specific models over one shared tokenizer, all Apache 2.0:

• MOSS-TTSD-v1.0 (8B) — multi-speaker, ultra-long spoken dialogue; the model OpenMOSS says outperformed Doubao and Gemini 2.5-pro in subjective tests.
• MOSS-VoiceGenerator (labelled 1.7B; 2.11B by index) — voice design from a text prompt, with no reference audio at all.
• MOSS-TTS-Realtime (labelled 1.7B; 2.33B by index) — streaming, multi-turn context-aware synthesis for voice agents; claimed 180 ms TTFB.
• MOSS-TTS-Nano-100M (0.1B) — 48 kHz two-channel output, streaming generation "can run on a 4-core CPU," 20 languages.
• MOSS-SoundEffect (8B) — environmental and foley sound generation with controllable duration.
• MOSS-Audio-Tokenizer — the shared discrete audio interface; 343,719 downloads in 30 days on its own.
• OpenMOSS/MOSS-TTS on GitHub — inference code, LoRA scripts, per-architecture model cards; 3,742 stars as of July 8, 2026. The repository was created on February 7, 2026 and its first public commit, Init, landed on February 10.

Beyond speech, the same lab shipped MOVA on January 28, 2026 — an Apache 2.0 mixture-of-experts model the card describes as having "32B total parameters (18B active during inference)" that generates video and synchronized audio in a single inference pass, rather than dubbing a silent clip afterwards. OpenMOSS pitches it explicitly against closed systems: "In a field dominated by closed-source models (Sora 2, Veo 3, Kling), we are releasing model weights, inference code, training pipelines, and LoRA fine-tuning scripts." MOVA-360p reports 98,428 downloads in 30 days. Its paper, arXiv 2602.08794, was submitted on February 9, 2026 — twelve days after the weights appeared — and is credited to the "SII-OpenMOSS Team," further corroborating the SII reporting line. No independent ranking of open joint video-generation and audio models was found, so this page makes no download-leadership claim for MOVA.

Community & Resources

• OpenMOSS-Team/MOSS-TTS on Hugging Face — weights, model card, Seed-TTS-eval table
• OpenMOSS-Team/MOSS-TTS-v1.5 — the current checkpoint, May 25, 2026
• MOSS-TTS Technical Report — arXiv 2603.18090, submitted March 18, 2026, revised March 20
• OpenMOSS/MOSS-TTS on GitHub — Apache 2.0 inference code
• OpenMOSS on GitHub — "a research group under the Shanghai Innovation Institution (SII), working in close collaboration with Fudan University and MOSI Intelligence"
• MOVA: Towards Scalable and Synchronized Video-Audio Generation — the lab's joint video-audio model
• Artificial Analysis Speech Arena — where MOSS-TTS is, notably, not listed
• Compare with Qwen3.7-Max, MiniMax-M3, Z-Image, and Stable Diffusion 3.5