You can deploy this model and start earning money today!

OpenMOSS-Team/
MOSS-TTS-Local-Transformer

Name: MOSS-TTS-Local-Transformer
Rating: 2.5 (6198 reviews)

Views

No views yet

MOSS-TTS Family

Overview

MOSS‑TTS Family is an open‑source speech and sound generation model family from MOSI.AI and the OpenMOSS team. It is designed for high‑fidelity, high‑expressiveness, and complex real‑world scenarios, covering stable long‑form speech, multi‑speaker dialogue, voice/character design, environmental sound effects, and real‑time streaming TTS.

Introduction

When a single piece of audio needs to sound like a real person, pronounce every word accurately, switch speaking styles across content, remain stable over tens of minutes, and support dialogue, role‑play, and real‑time interaction, a single TTS model is often not enough. The MOSS‑TTS Family breaks the workflow into five production‑ready models that can be used independently or composed into a complete pipeline.

MOSS‑TTS: The flagship production model featuring high fidelity and optimal zero-shot voice cloning. It supports long-speech generation, fine-grained control over Pinyin, phonemes, and duration, as well as multilingual/code-switched synthesis.
MOSS‑TTSD: A spoken dialogue generation model for expressive, multi-speaker, and ultra-long dialogues. The new v1.0 version achieves industry-leading performance on objective metrics and outperformed top closed-source models like Doubao and Gemini 2.5-pro in subjective evaluations. You can visit the MOSS-TTSD repository for details.
MOSS‑VoiceGenerator: An open-source voice design model capable of generating diverse voices and styles directly from text prompts, without any reference speech. It unifies voice design, style control, and synthesis, functioning independently or as a design layer for downstream TTS. Its performance surpasses other top-tier voice design models in arena ratings.
MOSS‑TTS‑Realtime: A multi-turn context-aware model for real-time voice agents. It uses incremental synthesis to ensure natural and coherent replies, making it ideal for building low-latency voice agents when paired with text models. The TTFB (Time To First Byte) of MOSS-TTS-Realtime reaches 180 ms, and the $T_{\text{LLM-first-sentence}} + T_{\text{MOSS-TTS-Realtime-TTFB}}$ is 377 ms.
MOSS‑SoundEffect: A content creation model specialized in sound effect generation with wide category coverage and controllable duration. It generates audio for natural environments, urban scenes, biological sounds, human actions, and musical fragments, suitable for film, games, and interactive experiences.

Model Architecture

We train MossTTSDelay and MossTTSLocal as complementary baselines under one training/evaluation setup: Delay emphasizes long-context stability, inference speed, and production readiness, while Local emphasizes lightweight flexibility and strong objective performance for streaming-oriented systems. Together they provide reproducible references for deployment and research.

MossTTSRealtime is not a third comparison baseline; it is a capability-driven design for voice agents. By modeling multi-turn context from both prior text and user acoustics, it delivers low-latency streaming speech that stays coherent and voice-consistent across turns.

Architecture	Core Mechanism	Arch Details
`MossTTSDelay`	Multi‑head parallel RVQ prediction with delay‑pattern scheduling	Arch Details
`MossTTSLocal`	Time‑synchronous RVQ blocks with a depth transformer	Arch Details
`MossTTSRealtime`	Hierarchical text–audio inputs for realtime synthesis	Arch Details

Released Models

Model	Architecture	Size	Model Card	Hugging Face	ModelScope
MOSS-TTS	`MossTTSDelay`	8B	Model Card	Hugging Face	ModelScope
	`MossTTSLocal`	1.7B	Model Card	Hugging Face	ModelScope
MOSS‑TTSD‑V1.0	`MossTTSDelay`	8B	Model Card	Hugging Face	ModelScope
MOSS‑VoiceGenerator	`MossTTSDelay`	1.7B	Model Card	Hugging Face	ModelScope
MOSS‑SoundEffect	`MossTTSDelay`	8B	Model Card	Hugging Face	ModelScope
MOSS‑TTS‑Realtime	`MossTTSRealtime`	1.7B	Model Card	Hugging Face	ModelScope

Supported Languages

MOSS-TTS, MOSS-TTSD and MOSS-TTS-Realtime currently supports 20 languages:

Language	Code	Flag	Language	Code	Flag	Language	Code	Flag
Chinese	zh	🇨🇳	English	en	🇺🇸	German	de	🇩🇪
Spanish	es	🇪🇸	French	fr	🇫🇷	Japanese	ja	🇯🇵
Italian	it	🇮🇹	Hungarian	hu	🇭🇺	Korean	ko	🇰🇷
Russian	ru	🇷🇺	Persian (Farsi)	fa	🇮🇷	Arabic	ar	🇸🇦
Polish	pl	🇵🇱	Portuguese	pt	🇵🇹	Czech	cs	🇨🇿
Danish	da	🇩🇰	Swedish	sv	🇸🇪
Greek	el	🇬🇷	Turkish	tr	🇹🇷

MOSS-TTS

1. Overview

1.1 TTS Family Positioning

MOSS-TTS is the flagship base model in our open-source TTS Family. It is designed as a production-ready synthesis backbone that can serve as the primary high-quality engine for scalable voice applications, and as a strong research baseline for controllable TTS and discrete audio token modeling.

Design goals

Production readiness: robust voice cloning with stable, on-brand speaker identity at scale
Controllability: duration and pronunciation controls that integrate into real workflows
Long-form stability: consistent identity and delivery for extended narration
Multilingual coverage: multilingual and code-switched synthesis as first-class capabilities

1.2 Key Capabilities

MOSS-TTS delivers state-of-the-art quality while providing the fine-grained controllability and long-form stability required for production-grade voice applications, from zero-shot cloning and hour-long narration to token- and phoneme-level control across multilingual and code-switched speech.

State-of-the-art evaluation performance — top-tier objective and subjective results across standard TTS benchmarks and in-house human preference testing, validating both fidelity and naturalness.
Zero-shot Voice Cloning (Voice Clone) — clone a target speaker’s timbre (and part of speaking style) from short reference audio, without speaker-specific fine-tuning.
Ultra-long Speech Generation (up to 1 hour) — support continuous long-form speech generation for up to one hour in a single run, designed for extended narration and long-session content creation.
Token-level Duration Control — control pacing, rhythm, pauses, and speaking rate at token resolution for precise alignment and expressive delivery.
Phoneme-level Pronunciation Control — supports:
- pure Pinyin input
- pure IPA phoneme input
- mixed Chinese / English / Pinyin / IPA input in any combination
Multilingual support — high-quality multilingual synthesis with robust generalization across languages and accents.
Code-switching — natural mixed-language generation within a single utterance (e.g., Chinese–English), with smooth transitions, consistent speaker identity, and pronunciation-aware rendering on both sides of the switch.

1.3 Model Architecture

MOSS-TTS includes two complementary architectures, both trained and released to explore different performance/latency tradeoffs and to support downstream research.

Architecture A: Delay Pattern (MossTTSDelay)

Single Transformer backbone with (n_vq + 1) heads.
Uses delay scheduling for multi-codebook audio tokens.
Strong long-context stability, efficient inference, and production-friendly behavior.

Architecture B: Global Latent + Local Transformer (MossTTSLocal)

Backbone produces a global latent per time step.
A lightweight Local Transformer emits a token block per step.
Streaming-friendly with simpler alignment (no delay scheduling).

Why train both?

Exploration of architectural potential and validation across multiple generation paradigms.
Different tradeoffs: Delay pattern tends to be faster and more stable for long-form synthesis; Local is smaller and excels on objective benchmarks.
Open-source value: two strong baselines for research, ablation, and downstream innovation.

For full details, see:

moss_tts_delay/README.md
moss_tts_local/README.md

1.4 Released Models

Model	Description
MossTTSDelay-8B	Recommended for production. Faster inference, stronger long-context stability, and robust voice cloning quality. Best for large-scale deployment and long-form narration.
MossTTSLocal-1.7B	Recommended for evaluation and research. Smaller model size with SOTA objective metrics. Great for quick experiments, ablations, and academic studies.

Recommended decoding hyperparameters (per model)

Model	audio_temperature	audio_top_p	audio_top_k	audio_repetition_penalty
MOSS-TTSDelay-8B	1.7	0.8	25	1.0
MOSS-TTSLocal-1.7B	1.0	0.95	50	1.1

Note: max_new_tokens controls duration. At 12.5 tokens per second, 1s ≈ 12.5 tokens.

2. Quick Start

Environment Setup

We recommend a clean, isolated Python environment with Transformers 5.0.0 to avoid dependency conflicts.

bash

conda create -n moss-tts python=3.12 -y
conda activate moss-tts

Install all required dependencies:

bash

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

(Optional) Install FlashAttention 2

For better speed and lower GPU memory usage, you can install FlashAttention 2 if your hardware supports it.

pip install --extra-index-url https://download.pytorch.org/whl/cu128 -e ".[flash-attn]"

If your machine has limited RAM and many CPU cores, you can cap build parallelism:

MAX_JOBS=4 pip install --extra-index-url https://download.pytorch.org/whl/cu128 -e ".[flash-attn]"

Notes:

Dependencies are managed in pyproject.toml, which currently pins torch==2.9.1+cu128 and torchaudio==2.9.1+cu128.
If FlashAttention 2 fails to build on your machine, you can skip it and use the default attention backend.
FlashAttention 2 is only available on supported GPUs and is typically used with torch.float16 or torch.bfloat16.

Basic Usage

Tip: For evaluation and research purposes, we recommend using MOSS-TTSLocal-1.7B.

MOSS-TTS provides a convenient generate interface for rapid usage. The examples below cover:

Direct generation (Chinese / English / Pinyin / IPA)
Voice cloning
Duration control

python

import importlib.util
from pathlib import Path
import torch
import torchaudio
from transformers import AutoModel, AutoProcessor, GenerationConfig
# Disable the broken cuDNN SDPA backend
torch.backends.cuda.enable_cudnn_sdp(False)
# Keep these enabled as fallbacks
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_mem_efficient_sdp(True)
torch.backends.cuda.enable_math_sdp(True)

pretrained_model_name_or_path = "OpenMOSS-Team/MOSS-TTS-Local-Transformer"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

def resolve_attn_implementation() -> str:
    # Prefer FlashAttention 2 when package + device conditions are met.
    if (
        device == "cuda"
        and importlib.util.find_spec("flash_attn") is not None
        and dtype in {torch.float16, torch.bfloat16}
    ):
        major, _ = torch.cuda.get_device_capability()
        if major >= 8:
            return "flash_attention_2"

    # CUDA fallback: use PyTorch SDPA kernels.
    if device == "cuda":
        return "sdpa"

    # CPU fallback.
    return "eager"


attn_implementation = resolve_attn_implementation()
print(f"[INFO] Using attn_implementation={attn_implementation}")

processor = AutoProcessor.from_pretrained(
    pretrained_model_name_or_path,
    trust_remote_code=True,
)
processor.audio_tokenizer = processor.audio_tokenizer.to(device)

text_1 = """亲爱的你，
你好呀。

今天，我想用最认真、最温柔的声音，对你说一些重要的话。
这些话，像一颗小小的星星，希望能在你的心里慢慢发光。

首先，我想祝你——
每天都能平平安安、快快乐乐。

希望你早上醒来的时候，
窗外有光，屋子里很安静，
你的心是轻轻的，没有着急，也没有害怕。
"""
text_2 = """We stand on the threshold of the AI era.
Artificial intelligence is no longer just a concept in laboratories, but is entering every industry, every creative endeavor, and every decision. It has learned to see, hear, speak, and think, and is beginning to become an extension of human capabilities. AI is not about replacing humans, but about amplifying human creativity, making knowledge more equitable, more efficient, and allowing imagination to reach further. A new era, jointly shaped by humans and intelligent systems, has arrived."""
text_3 = "nin2 hao3，qing3 wen4 nin2 lai2 zi4 na3 zuo4 cheng2 shi4？"
text_4 = "nin2 hao3，qing4 wen3 nin2 lai2 zi4 na4 zuo3 cheng4 shi3？"
text_5 = "您好，请问您来自哪 zuo4 cheng2 shi4？"
text_6 = "/həloʊ, meɪ aɪ æsk wɪtʃ sɪti juː ɑːr frʌm?/"

# Use audio from ./assets/audio to avoid downloading from the cloud.
ref_audio_1 = "https://speech-demo.oss-cn-shanghai.aliyuncs.com/moss_tts_demo/tts_readme_demo/reference_zh.wav"
ref_audio_2 = "https://speech-demo.oss-cn-shanghai.aliyuncs.com/moss_tts_demo/tts_readme_demo/reference_en.m4a"

conversations = [
    # Direct TTS (no reference)
    [
        processor.build_user_message(text=text_1)
    ],
    [
        processor.build_user_message(text=text_2)
    ],
    # Pinyin or IPA input
    [
        processor.build_user_message(text=text_3)
    ],
    [
        processor.build_user_message(text=text_4)
    ],
    [
        processor.build_user_message(text=text_5)
    ],
    [
        processor.build_user_message(text=text_6)
    ],
    # Voice cloning (with reference)
    [
        processor.build_user_message(text=text_1, reference=[ref_audio_1])
    ],
    [
        processor.build_user_message(text=text_2, reference=[ref_audio_2])
    ],
]

model = AutoModel.from_pretrained(
    pretrained_model_name_or_path,
    trust_remote_code=True,
    attn_implementation=attn_implementation,
    torch_dtype=dtype,
).to(device)
model.eval()

batch_size = 1

save_dir = Path(f"inference_root_moss_tts_local_transformer_generation")
save_dir.mkdir(exist_ok=True, parents=True)
sample_idx = 0
with torch.no_grad():
    for start in range(0, len(conversations), batch_size):
        batch_conversations = conversations[start : start + batch_size]
        batch = processor(batch_conversations, mode="generation")
        input_ids = batch["input_ids"].to(device)
        attention_mask = batch["attention_mask"].to(device)

        outputs = model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            max_new_tokens=4096,
        )

        for message in processor.decode(outputs):
            audio = message.audio_codes_list[0]
            out_path = save_dir / f"sample{sample_idx}.wav"
            sample_idx += 1
            torchaudio.save(out_path, audio.unsqueeze(0), processor.model_config.sampling_rate)

Continuation + Voice Cloning (Prefix Audio + Text)

MOSS-TTS supports continuation-based cloning: provide a prefix audio clip in the assistant message, and make sure the prefix transcript is included in the text. The model continues in the same speaker identity and style.

python

import importlib.util
from pathlib import Path
import torch
import torchaudio
from transformers import AutoModel, AutoProcessor, GenerationConfig
# Disable the broken cuDNN SDPA backend
torch.backends.cuda.enable_cudnn_sdp(False)
# Keep these enabled as fallbacks
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_mem_efficient_sdp(True)
torch.backends.cuda.enable_math_sdp(True)

pretrained_model_name_or_path = "OpenMOSS-Team/MOSS-TTS-Local-Transformer"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

def resolve_attn_implementation() -> str:
    # Prefer FlashAttention 2 when package + device conditions are met.
    if (
        device == "cuda"
        and importlib.util.find_spec("flash_attn") is not None
        and dtype in {torch.float16, torch.bfloat16}
    ):
        major, _ = torch.cuda.get_device_capability()
        if major >= 8:
            return "flash_attention_2"

    # CUDA fallback: use PyTorch SDPA kernels.
    if device == "cuda":
        return "sdpa"

    # CPU fallback.
    return "eager"


attn_implementation = resolve_attn_implementation()
print(f"[INFO] Using attn_implementation={attn_implementation}")

processor = AutoProcessor.from_pretrained(
    pretrained_model_name_or_path,
    trust_remote_code=True,
)
processor.audio_tokenizer = processor.audio_tokenizer.to(device)

text_1 = """亲爱的你，
你好呀。

今天，我想用最认真、最温柔的声音，对你说一些重要的话。
这些话，像一颗小小的星星，希望能在你的心里慢慢发光。

首先，我想祝你——
每天都能平平安安、快快乐乐。

希望你早上醒来的时候，
窗外有光，屋子里很安静，
你的心是轻轻的，没有着急，也没有害怕。
"""

ref_text_1 = "太阳系八大行星之一。"
# Use audio from ./assets/audio to avoid downloading from the cloud.
ref_audio_1 = "https://speech-demo.oss-cn-shanghai.aliyuncs.com/moss_tts_demo/tts_readme_demo/reference_zh.wav"

conversations = [
    # Continuatoin only
    [
        processor.build_user_message(text=ref_text_1 + text_1),
        processor.build_assistant_message(audio_codes_list=[ref_audio_1])
    ],
]

model = AutoModel.from_pretrained(
    pretrained_model_name_or_path,
    trust_remote_code=True,
    attn_implementation=attn_implementation,
    torch_dtype=dtype,
).to(device)
model.eval()

batch_size = 1

save_dir = Path("inference_root_moss_tts_local_transformer_continuation")
save_dir.mkdir(exist_ok=True, parents=True)
sample_idx = 0
with torch.no_grad():
    for start in range(0, len(conversations), batch_size):
        batch_conversations = conversations[start : start + batch_size]
        batch = processor(batch_conversations, mode="continuation")
        input_ids = batch["input_ids"].to(device)
        attention_mask = batch["attention_mask"].to(device)

        outputs = model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            max_new_tokens=4096,
        )

        for message in processor.decode(outputs):
            audio = message.audio_codes_list[0]
            out_path = save_dir / f"sample{sample_idx}.wav"
            sample_idx += 1
            torchaudio.save(out_path, audio.unsqueeze(0), processor.model_config.sampling_rate)

Input Types

UserMessage

Field	Type	Required	Description
`text`	`str`	Yes	Text to synthesize. Supports Chinese, English, German, French, Spanish, Japanese, Korean, etc. Can mix raw text with Pinyin or IPA for pronunciation control.
`reference`	`List[str]`	No	Reference audio for voice cloning. For current MOSS-TTS, one audio is expected in the list.
`tokens`	`int`	No	Expected number of audio tokens. 1s ≈ 12.5 tokens.

AssistantMessage

Field	Type	Required	Description
`audio_codes_list`	`List[str]`	Only for continuation	Prefix audio for continuation-based cloning. Use audio file paths or URLs.

Generation Hyperparameters (MOSS-TTS-Local)

MOSS-TTSLocal utilizes DelayGenerationConfig to manage hierarchical sampling. Due to the Progressive Sequence Dropout training mechanism, the model supports variable bitrate inference by adjusting the RVQ depth.

Parameter	Type	Recommended (Audio Layers)	Description
`max_new_tokens`	`int`	—	Controls total generated audio tokens. 1s ≈ 12.5 tokens.
`n_vq_for_inference`	`int`	32	RVQ Inference Depth: Controls the number of codebook layers generated. Higher values (max 32) improve audio fidelity but slow down inference; lower values speed up inference but reduce audio quality.
`audio_temperature`	`float`	1.0	Temperature for audio token layers (Layer 1+). Lower values ensure more stable and consistent acoustic reconstruction.
`audio_top_p`	`float`	0.95	Nucleus sampling cutoff for audio layers.
`audio_top_k`	`int`	50	Top-K sampling filter for audio layers.
`audio_repetition_penalty`	`float`	1.1	Discourages repeating acoustic patterns. Values > 1.0 help prevent artifacts in long-form synthesis.

Pinyin Input

Use tone-numbered Pinyin such as ni3 hao3 wo3 men1. You can convert Chinese text with pypinyin, then adjust tones for pronunciation control.

python

import re
from pypinyin import pinyin, Style

CN_PUNCT = r"，。！？；：、（）“”‘’"


def fix_punctuation_spacing(s: str) -> str:
    s = re.sub(rf"\s+([{CN_PUNCT}])", r"\1", s)
    s = re.sub(rf"([{CN_PUNCT}])\s+", r"\1", s)
    return s


def zh_to_pinyin_tone3(text: str, strict: bool = True) -> str:
    result = pinyin(
        text,
        style=Style.TONE3,
        heteronym=False,
        strict=strict,
        errors="default",
    )

    s = " ".join(item[0] for item in result)
    return fix_punctuation_spacing(s)

text = zh_to_pinyin_tone3("您好，请问您来自哪座城市？")
print(text)

# Expected: nin2 hao3，qing3 wen4 nin2 lai2 zi4 na3 zuo4 cheng2 shi4？
# Try: nin2 hao3，qing4 wen3 nin2 lai2 zi4 na4 zuo3 cheng4 shi3？

IPA Input

Use /.../ to wrap IPA sequences so they are distinct from normal text. You can use DeepPhonemizer to convert English paragraphs or words into IPA sequences.

python

from dp.phonemizer import Phonemizer

# Download a phonemizer checkpoint from https://public-asai-dl-models.s3.eu-central-1.amazonaws.com/DeepPhonemizer/en_us_cmudict_ipa_forward.pt
model_path = "<path-to-phonemizer-checkpoint>"
phonemizer = Phonemizer.from_checkpoint(model_path)

english_texts = "Hello, may I ask which city you are from?"
phoneme_outputs = phonemizer(
    english_texts,
    lang="en_us",
    batch_size=8
)
model_input_text = f"/{phoneme_outputs}/"
print(model_input_text)

# Expected: /həloʊ, meɪ aɪ æsk wɪtʃ sɪti juː ɑːr frʌm?/

3. Evaluation

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.

Model	Params	Open‑source	EN WER (%) ↓	EN SIM (%) ↑	ZH CER (%) ↓	ZH SIM (%) ↑
DiTAR	0.6B	❌	1.69	73.5	1.02	75.3
FishAudio‑S1	4B	❌	1.72	62.57	1.22	72.1
CosyVoice3	1.5B	❌	2.22	72	1.12	78.1
Seed‑TTS		❌	2.25	76.2	1.12	79.6
MiniMax‑Speech		❌	1.65	69.2	0.83	78.3

CosyVoice	0.3B	✅	4.29	60.9	3.63	72.3
CosyVoice2	0.5B	✅	3.09	65.9	1.38	75.7
CosyVoice3	0.5B	✅	2.02	71.8	1.16	78
F5‑TTS	0.3B	✅	2	67	1.53	76
SparkTTS	0.5B	✅	3.14	57.3	1.54	66
FireRedTTS	0.5B	✅	3.82	46	1.51	63.5
FireRedTTS‑2	1.5B	✅	1.95	66.5	1.14	73.6
Qwen2.5‑Omni	7B	✅	2.72	63.2	1.7	75.2
FishAudio‑S1‑mini	0.5B	✅	1.94	55	1.18	68.5
IndexTTS2	1.5B	✅	2.23	70.6	1.03	76.5
VibeVoice	1.5B	✅	3.04	68.9	1.16	74.4
HiggsAudio‑v2	3B	✅	2.44	67.7	1.5	74
GLM-TTS	1.5B	✅	2.23	67.2	1.03	76.1
GLM-TTS-RL	1.5B	✅	1.91	68.1	0.89	76.4
VoxCPM	0.5B	✅	1.85	72.9	0.93	77.2
Qwen3‑TTS	0.6B	✅	1.68	70.39	1.23	76.4
Qwen3‑TTS	1.7B	✅	1.5	71.45	1.33	76.72

MossTTSDelay	8B	✅	1.84	70.86	1.37	76.98
MossTTSLocal	1.7B	✅	1.93	73.28	1.44	79.62